Radiomics and anaesthetic planning: Quantitative imaging as a new frontier in preoperative risk assessment

Abstract

Radiomics-the high-throughput extraction of quantitative features from standard medical images-has transformed oncologic imaging by revealing subvisual patterns linked to tissue biology, yet its role in perioperative medicine remains largely unexplored. This narrative review summarises current evidence linking imaging-derived radiomic biomarkers to perioperative outcomes and proposes a conceptual framework for integrating radiomics into precision anaesthesia. Quantitative assessment of body composition, organ function, and vascular morphology from routine preoperative computed tomography and magnetic resonance imaging can provide objective indicators of physiologic reserve, drug-handling capacity, and recovery potential. Across heterogeneous, largely oncological cohorts, combined radiomic-clinical models have reported higher discrimination (area under the curve 0.84-0.93) than conventional risk scores for selected postoperative complications, and artificial intelligence-based airway assessment has shown sensitivity and specificity exceeding traditional bedside tests; these figures are pooled from methodologically diverse studies rather than single validated estimates. The Image Biomarker Standardisation Initiative has substantially reduced cross-platform feature variability. However, perioperative-specific, prospectively validated evidence remains scarce. Prospective multicentre trials, standardized and automated feature-extraction pipelines, transparent cost and equity appraisal, and integration with electronic health records are critical priorities before radiomics-driven preoperative assessment can enter routine anaesthetic practice.

Key Words: Radiomics; Anaesthesia; Perioperative medicine; Preoperative risk assessment; Body composition; Sarcopenia; Machine learning; Precision medicine; Quantitative imaging; Image biomarkers

Core Tip: This minireview proposes radiomics-the extraction of high-dimensional quantitative features from routine medical images-as a novel tool for precision anaesthesia. By quantifying body composition, organ function, and vascular morphology from preoperative computed tomography and magnetic resonance imaging, clinicians can obtain objective biomarkers of physiologic reserve that augment traditional preoperative evaluation. Integration of radiomic features with clinical data through machine learning may enable personalized anaesthetic planning, optimized drug dosing, and improved identification of high-risk surgical patients.

INTRODUCTION

Preoperative risk assessment remains a cornerstone of safe anaesthetic practice. Conventional approaches rely on clinical history, physical examination, laboratory investigations, and validated scoring systems such as the American Society of Anesthesiologists (ASA) Physical Status classification, the Revised Cardiac Risk Index, and the Surgical Outcome Risk Tool[1,2]. While these instruments provide valuable prognostic information, they are inherently limited by their reliance on subjective clinical judgement, categorical variables, and population-level risk estimates that may inadequately capture individual physiologic reserve[3].

Simultaneously, the field of medical imaging has undergone a paradigm shift. The emergence of radiomics-defined by Lambin et al[4] as the high-throughput extraction of large amounts of quantitative features from radiographic images-has revolutionized the analysis of routine diagnostic imaging. First formally described in 2012[4] and subsequently conceptualized as a bridge between medical imaging and personalized medicine[5], radiomics converts standard computed tomography (CT), magnetic resonance imaging (MRI), and positron emission tomography images into mineable, high-dimensional data. As Gillies et al[6] articulated in their seminal 2016 review, “images are more than pictures, they are data”, underscoring the vast untapped information reservoir within clinical imaging studies.

The landmark study by Aerts et al[7] in 2014, which extracted 440 radiomic features from CT data of 1019 patients with lung or head-and-neck cancer, demonstrated that quantitative image features possess prognostic power independent of clinical variables and correlate with underlying gene-expression patterns. This work catalyzed an exponential growth in radiomics research, with annual publications increasing from approximately 120 in 2017 to over 1500 in 2023[8].

While radiomics has been predominantly applied in oncology for tumour characterisation, treatment response prediction, and prognostication, its potential applications in perioperative medicine and anaesthetic planning remain largely unexplored. This represents a significant missed opportunity, as virtually every surgical patient undergoes preoperative imaging that contains quantitative information about body composition, organ function, and vascular status-all critical determinants of anaesthetic risk and postoperative outcome.

The concept of precision anaesthesia-the personalized integration of a patient’s genetic blueprint, clinical history, and physiologic parameters to tailor anaesthetic care-has gained momentum in recent years[9]. Pharmacogenomic studies have demonstrated that polymorphisms in cytochrome P450 enzymes (CYP2B6, CYP3A4, CYP2D6) significantly influence the metabolism of propofol, midazolam, fentanyl, and other anaesthetic agents[9]. However, the contribution of quantitative imaging biomarkers to this precision paradigm has received minimal attention.

This narrative review aims to: (1) Introduce the radiomics workflow and feature taxonomy relevant to perioperative assessment; (2) Summarise current evidence linking imaging-derived biomarkers of body composition, organ function, and vascular morphology to anaesthetic outcomes; (3) Evaluate emerging applications of artificial intelligence (AI) in airway assessment and frailty quantification; (4) Discuss challenges in standardisation and reproducibility; and (5) Outline research priorities for clinical translation of radiomics-driven preoperative evaluation.

Before surveying this evidence, it is important to define precisely what radiomics uniquely contributes, and what it does not. Conventional multimodal preoperative assessment-combining clinical risk scores, pulmonary function tests, echocardiography, and the qualitative radiologist’s report-already captures substantial prognostic information. The distinct, non-redundant value proposition of radiomics is fourfold: First, it quantifies operator-independent, tissue-level information (texture, heterogeneity, and spatial pattern) that is imperceptible to the human eye and absent from the categorical radiological report; second, it derives this information from imaging already obtained for surgical planning, imposing no additional radiation, cost, or patient burden; third, it permits simultaneous phenotyping of multiple organ systems from a single acquisition; and fourth, it yields continuous, reproducible biomarkers that can be embedded directly into machine-learning risk models, in contrast to the coarse categorical outputs of traditional scores. Conversely, radiomics does not replace definitive functional testing, does not yet provide causal or decision-grade drug-dosing rules, and offers limited incremental value where a direct functional measurement is already available. This review is therefore organized around that critically defined thesis rather than as an exhaustive catalogue of applications.

LITERATURE SEARCH AND STUDY SELECTION

This article is a narrative review. A structured literature search was performed in four databases-PubMed, Scopus, Web of Science, and Google Scholar-covering January 2012 (the year radiomics was formally described) to February 2026. The database-specific search strings, field tags, and date limits are summarized in Table 1.

Table 1 Search strategy used for the narrative review.

Database	Search string/field tags	Date range	Limits/filters
PubMed	[Radiomic* (tiab) OR radiomics (tiab) OR “quantitative imaging”(tiab) OR “texture analysis”(tiab)] AND [CT (tiab) OR “computed tomography”(tiab) OR MRI (tiab) OR “magnetic resonance imaging”(tiab) OR PET (tiab)] AND [body composition (tiab) OR sarcopenia(tiab) OR myosteatosis (tiab) OR visceral adiposity (tiab) OR liver (tiab) OR cardiac (tiab) OR pulmonary (tiab) OR renal (tiab) OR vascular (tiab) OR airway (tiab)] AND [perioperative (tiab) OR postoperative (tiab) OR surgical outcome (tiab) OR anaesthe* (tiab) OR anesthe* (tiab) OR “risk prediction” (tiab) OR “machine learning” (tiab)]	2012/01/01-2026/02/28	English language; humans
Scopus	TITLE-ABS-KEY (radiomic* OR “quantitative imaging”) AND TITLE-ABS-KEY (CT OR MRI OR PET) AND TITLE-ABS-KEY (“body composition” OR sarcopenia OR airway OR vascular OR cardiac OR pulmonary OR liver OR renal) AND TITLE-ABS-KEY (perioperative OR postoperative OR anaesthesia OR anesthesia OR “machine learning”)	2012-2026	English; article, review
Web of Science	TS = (radiomic* OR “quantitative imaging”) AND TS = (CT OR MRI OR PET) AND TS = (“body composition” OR sarcopenia OR airway OR vascular OR cardiac OR pulmonary OR liver OR renal) AND TS = (perioperative OR postoperative OR anaesthesia OR anesthesia OR “machine learning”)	2012-2026	English; core collection
Google scholar	Radiomics OR “quantitative imaging” AND perioperative OR anaesthesia OR anesthesia AND “body composition” OR sarcopenia OR airway OR vascular	2012-2026	Title/keyword search; first 200 results screened by relevance

Studies were eligible for inclusion if they: (1) Reported radiomic or quantitative-imaging feature extraction from CT, MRI, or positron emission tomography; (2) Addressed body composition, organ-specific functional imaging, vascular morphology, airway assessment, or predictive modelling of surgical or perioperative outcomes; (3) Were original research articles, systematic reviews, or meta-analyses; and (4) Were published in English. Studies were excluded if they: (1) Addressed purely oncological endpoints (e.g., tumour grading or treatment-response prediction) with no extractable perioperative or physiologic-reserve relevance; (2) Were conference abstracts or non-peer-reviewed preprints, or editorials without primary data (except where a perspective was specifically pertinent to implementation); or (3) Provided insufficient methodological detail to judge feature-extraction validity. Where multiple reports addressed the same application, priority was given to larger, more recent, and externally validated studies.

The methodological rigour of the included radiomics studies was appraised narratively against the domains of the Radiomics Quality Score (RQS) proposed by Lambin et al[5] encompassing imaging-protocol documentation, segmentation reproducibility, feature robustness, internal and external validation, and data openness. A formal, item-by-item RQS score was not assigned to every primary study, and formal risk-of-bias instruments (QUADAS-2, PROBAST) and quantitative meta-analysis were not applied. This is a deliberate consequence of the narrative design: The included literature is highly heterogeneous in imaging modality, anatomical target, outcome definition, and reference standard, which precludes meaningful statistical pooling, and the purpose of the review is to map an emerging, cross-disciplinary field and generate hypotheses rather than to derive a pooled effect estimate. The evidence summarized here should accordingly be interpreted as exploratory and hypothesis-generating, not as a graded synthesis of diagnostic accuracy. A study-selection flow diagram is provided as Supplementary Figure 1. Counts for Scopus, Web of Science, and Google Scholar are reported as identified by the corresponding database searches; the full screening flow is detailed in Supplementary Figure 1.

Search performed January 2012 to February 2026. The PubMed core query was executed; the Scopus, Web of Science, and Google Scholar strings are representative and require minor syntactic adaptation to each interface. A study-selection flow diagram is provided as Supplementary Figure 1.

RADIOMICS: CONCEPT, WORKFLOW, AND FEATURE TAXONOMY

The radiomics pipeline

The radiomics workflow comprises seven sequential steps: Dataset definition, image acquisition, preprocessing, segmentation, feature extraction, feature selection, and model development[8] (Figure 1). Each step demands rigorous standardisation to ensure reproducibility and clinical applicability.

Figure 1 The radiomics pipeline for perioperative anaesthetic assessment. CT: Computed tomography; MRI: Magnetic resonance imaging; PET: Positron emission tomography; ROI: Region of interest; IBSI: Image biomarker standardisation initiative; GLCM/GLRLM/GLSZM/GLDM/NGTDM: Texture feature matrices; PACS: Picture archiving and communication system; EHR: Electronic health record; PPCs: Postoperative pulmonary complications; AKI: Acute kidney injury; AUC: Area under the curve; CV: Coefficient of variation; LoG: Laplacian of Gaussian; LASSO: Least absolute shrinkage and selection operator; mRMR: Minimum redundancy maximum relevance; PCA: Principal component analysis; SVM: Support vector machine; CNN: Convolutional neural network; SHAP: SHapley Additive exPlanations.

Figure 1 The radiomics pipeline for perioperative anaesthetic assessment. From routine preoperative imaging to precision anaesthesia through standardised quantitative feature analysis. The figure has been simplified to emphasize the linear pipeline flow.

Image acquisition requires consistent modality, standardized protocols, and minimized external variables, as radiomic features are highly sensitive to variations in imaging parameters[8]. Preprocessing encompasses resampling (to standardize voxel spacing), bin width setting (typically 30-130 bins), and normalisation (to reconcile scanner differences, e.g., 1.5T vs 3.0T MRI)[8]. Segmentation defines the region of interest and may be performed manually, semi-automatically, or automatically, with three-dimensional volumetric analysis providing more comprehensive characterisation than two-dimensional approaches[8].

Feature categories

Radiomic features are systematically categorised according to the Image Biomarker Standardisation Initiative (IBSI) framework[10] and implemented in software platforms such as PyRadiomics (Table 2).

Table 2 Classification of radiomic features with definitions and perioperative relevance.

Feature category	Description	Key features	Perioperative relevance
First-order (histogram)	Statistical measures of voxel intensity distribution without spatial consideration	Mean, median, entropy, skewness, kurtosis, energy, uniformity	Tissue density characterization; hepatic steatosis grading; muscle quality assessment (myosteatosis)
Shape-based	Geometric properties of the segmented region	Volume, surface area, sphericity, compactness, elongation	Organ volumetry; tumour burden estimation; airway dimensional analysis
GLCM	Second-order texture features capturing spatial arrangements of pixel intensity pairs	Contrast, correlation, homogeneity, energy, entropy	Tissue heterogeneity; myocardial fibrosis detection; pulmonary parenchymal characterization
GLRLM	Texture features based on consecutive pixels of identical intensity	Run length non-uniformity, short/long run emphasis, gray-level non-uniformity	Vascular calcification pattern analysis; skeletal muscle architecture
GLSZM	Features characterizing zones of homogeneous intensity	Zone entropy, large/small zone emphasis, zone non-uniformity	Emphysema quantification; adipose tissue distribution
GLDM	Features describing pixel-pair intensity dependencies	Dependence entropy, dependence non-uniformity	Coronary plaque vulnerability assessment
NGTDM	Intensity differences between adjacent pixels	Coarseness, busyness, complexity, contrast	Tissue microstructure characterization; perivascular adipose tissue analysis

GLCM/GLRLM/GLSZM/GLDM/NGTDM: Texture feature matrices.

First-order features demonstrate superior reproducibility compared to higher-order texture features[8], while second-order features (Gray Level Co-occurrence Matrix, Gray Level Run Length Matrix) capture spatial relationships that first-order statistics miss, providing critical information about tissue heterogeneity[11]. Optional image filters-including wavelet, Laplacian of Gaussian, and exponential transforms-can further enhance specific image characteristics prior to feature extraction[8].

BODY COMPOSITION ASSESSMENT: SARCOPENIA, ADIPOSITY, AND PHYSIOLOGIC RESERVE

CT-based sarcopenia quantification

Body composition is a fundamental determinant of anaesthetic pharmacokinetics, surgical risk, and postoperative recovery. Conventional assessment through body mass index (BMI) is a crude measure that fails to distinguish lean mass from adipose tissue or to detect sarcopenia-the pathologic loss of skeletal muscle mass and function[12]. CT-based body composition analysis offers superior diagnostic capability, detecting sarcopenia at a rate 27.3%-66.7% higher than BMI-based malnutrition screening[12].

The standard methodology involves measurement of skeletal muscle area at the third lumbar vertebra (L3) level on axial CT images, normalised to height squared to yield the Skeletal Muscle Index (SMI, cm²/m²)[12]. Widely adopted diagnostic thresholds include SMI < 52.4 cm²/m² for males and < 38.5 cm²/m² for females (Prado criteria), though population-specific cutoffs have been proposed[12]. Automated segmentation using U-Net three-dimensional neural networks achieves Sørensen-Dice coefficients of 0.955 with processing speeds under one second per scan, compared to approximately 15 minutes for manual analysis[12].

It must be emphasized, however, that the Prado thresholds were derived predominantly from North American (Western) oncology cohorts. Skeletal muscle mass, distribution, and overall body habitus differ substantially across ethnic groups, and several South and East Asian populations exhibit lower normative muscle mass; uniform application of Western cutoffs therefore risks both over-diagnosis and under-diagnosis of sarcopenia in non-Western populations. These thresholds must be prospectively validated, and ethnicity-specific cutoffs established, before any radiomics-derived sarcopenia threshold is universally integrated into preoperative screening algorithms. This limitation is revisited under equity and generalisability in the Challenges section.

The perioperative implications of CT-detected sarcopenia are profound. In a multicentre study of split liver transplantation recipients, postoperative worsening of sarcopenia and myosteatosis were identified as independent risk factors for mortality, with adverse effects on early allograft dysfunction, intensive care unit stay, and overall survival[13]. A systematic review and meta-analysis of patients with gastric cancer found that preoperative sarcopenia was associated with significantly higher rates of major postoperative complications [odds ratio (OR) 1.67] and pulmonary complications (OR 4.01) and with worse overall survival [hazard ratio (HR) 2.12][14]. For elderly hepatocellular carcinoma patients, sarcopenia measured by Psoas Muscle Index was associated with significantly increased mortality (HR 2.15, P = 0.032)[15].

Beyond mere mass quantification, CT radiomics of skeletal muscle provides additional prognostic information. Radiomic sarcopenia-defined using texture and heterogeneity features beyond simple cross-sectional area-correlated with both complications and long-term survival in gastric cancer patients, improving the accuracy of survival and complication prediction compared to conventional sarcopenia measures[16]. Machine learning models using CT radiomics of skeletal muscle achieved accuracy of 0.90 ± 0.03 for sarcopenia detection and 0.88 ± 0.04 for tumour progression prediction in gastric and oesophageal cancer patients[17].

Visceral adiposity and anaesthetic pharmacokinetics

Visceral adipose tissue (VAT) accumulation profoundly influences both surgical risk and anaesthetic drug behaviour. CT measurement at the L3 level using Hounsfield Unit ranges of -150 to -50 allows reliable segmentation of VAT from subcutaneous adipose tissue[18]. The visceral-to-subcutaneous fat ratio provides additional prognostic information, with VAT ≥ 100 cm² indicating clinically significant visceral obesity[12].

A 2025 study of 309 ovarian cancer patients demonstrated that visceral obesity (visceral fat area > 330.08 cm² at the T12 level) was an independent predictor of postoperative complications in younger patients (≤ 65 years), with complication rates of 56% vs 36% in non-obese counterparts (OR 1.98, P = 0.031)[18]. From an anaesthetic perspective, increased fat mass directly affects the volume of distribution of lipophilic drugs, including propofol, thiopental, and benzodiazepines, necessitating weight-based dosing adjustments[19].

The practical consequence can be illustrated with propofol. An induction dose scaled to total body weight tends to overestimate the dose required for the central compartment in a patient with high VAT, predisposing to hypotension and delayed emergence, whereas a dose scaled to ideal body weight may be insufficient because it ignores the enlarged peripheral fat reservoir. A radiomics-derived VAT score, by quantifying the actual lipophilic-drug distribution compartment, could in principle support a more physiologically grounded strategy-an induction dose titrated to lean (or adjusted) body weight, with the maintenance infusion rate informed by the size of the peripheral fat compartment-rather than reliance on total or ideal body weight alone. We emphasize that this remains a mechanistically plausible, hypothesis-generating application: No prospective study has yet correlated radiomic adiposity metrics with measured propofol pharmacokinetics, and such validation is a prerequisite before any radiomics-informed dosing rule can be recommended.

The integration of CT body composition analysis into preoperative anaesthetic assessment could therefore inform drug-dosing strategies, predict difficulty with regional anaesthetic techniques, and identify patients at elevated risk of postoperative complications (Table 3).

Table 3 Computed tomography-derived body composition parameters and their anaesthetic implications.

Parameter	Measurement method	Diagnostic threshold	Anaesthetic/perioperative significance
Skeletal Muscle Index	L3 axial CT area/height2	< 52.4 cm2/m2 (M), < 38.5 cm2/m2 (F)	Reduced physiologic reserve; prolonged recovery; increased ICU stay; altered drug metabolism
Skeletal muscle density	Mean HU of skeletal muscle at L3	< 41 HU (BMI < 25), < 33 HU (BMI ≥ 25)	Myosteatosis; worse perioperative morbidity and mortality than sarcopenia alone
Visceral adipose tissue	Area within peritoneal cavity at L3 (HU: -150 to -50)	≥ 100 cm2	Increased volume of distribution for lipophilic anaesthetics; technical difficulty with neuraxial/regional blocks
Subcutaneous adipose tissue	Area outside muscle fascia at L3	Context-dependent	Needle length considerations for regional anaesthesia; injection site planning
V/S ratio	VAT/SAT	> 0.4 (metabolic risk)	Metabolic syndrome; increased cardiovascular perioperative risk; insulin resistance
Psoas Muscle Index	Bilateral psoas area/height2	Population-specific	Frailty surrogate; postoperative mortality predictor (HR 2.15)

CT: Computed tomography; HU: Hounsfield units; BMI: Body mass index; ICU: Intensive care unit; M: Male; F: Female; L3: Third lumbar vertebra; VAT: Visceral adipose tissue; SAT: Subcutaneous adipose tissue; HR: Hazard ratio.

ORGAN-SPECIFIC RADIOMICS FOR PREOPERATIVE FUNCTIONAL ASSESSMENT

Hepatic assessment: Steatosis, fibrosis, and drug metabolism

The liver is the primary site of metabolism for the majority of anaesthetic agents. Hepatic steatosis-present in approximately 25% of the global population-alters drug clearance, increases sensitivity to hepatotoxins, and predisposes to perioperative liver injury[20]. Conventional CT assessment of hepatic steatosis through liver attenuation measurements provides moderate accuracy, with pooled sensitivity and specificity of 82% and 94%, respectively, for detecting at least moderate steatosis (> 20%-33% fat at biopsy)[20].

CT radiomics substantially enhances diagnostic capability beyond simple attenuation measurements. Clinical-radiomic models combining texture features extracted from non-contrast CT with clinical parameters have demonstrated accurate non-invasive prediction of nonalcoholic steatohepatitis[21]. A 2025 study demonstrated that two-dimensional radiomics models based on random forest and bagging decision tree algorithms showed high consistency with quantitative CT-based classification for diagnosing fatty liver[22]. Furthermore, CT radiomics can differentiate among diffuse liver diseases on non-contrast CT with clinically useful accuracy[23].

For the anaesthesiologist, preoperative identification of significant hepatic steatosis through radiomics enables: (1) Anticipation of altered drug metabolism and prolonged drug effects; (2) Avoidance of potentially hepatotoxic agents; (3) Risk stratification for postoperative liver failure in patients undergoing hepatic resection; and (4) Prediction of inadequate liver hypertrophy after portal vein embolization, where low muscle mass on CT body composition is an independent predictor of poor liver hypertrophy and post-hepatectomy liver failure[12].

Cardiac radiomics: Myocardial characterisation and perioperative risk

Cardiac CT radiomics has emerged as a rapidly expanding field; a recent systematic review and meta-analysis by the EuSoMII Radiomics Auditing Group included 202 studies of cardiac CT and MRI radiomics[24]. Applications span coronary plaque characterisation, perivascular adipose tissue analysis, myocardial tissue assessment, and intracardiac lesion classification.

Coronary plaque analysis: Kolossváry et al[25] demonstrated that, whereas none of the conventional quantitative CT metrics distinguished plaques exhibiting the napkin-ring sign-a marker of plaque vulnerability-from matched control plaques, a large proportion of radiomic features did, with the best-performing features achieving area under the curve (AUC) values of approximately 0.89-0.92.

Myocardial tissue characterisation: Hinzpeter et al[26] reported that CT texture features-specifically kurtosis and short-run high gray-level emphasis-achieved AUC of 0.90 (95% confidence interval: 0.80-0.99) for acute myocardial infarction prediction. Mannil et al[27] demonstrated sensitivity of 86% and specificity of 81% for myocardial pathology classification using machine learning on CT texture data. In hypertrophic cardiomyopathy, radiomics-based risk prediction yielded higher C-statistics than the European Society of Cardiology sudden cardiac death risk score[24]. A cardiac amyloidosis radiomics model achieved AUC values of 0.95, 0.95, and 0.91 in training, internal validation, and external validation cohorts, respectively[24].

Epicardial adipose tissue radiomics: Epicardial adipose tissue (EAT) radiomics has shown particular promise for perioperative cardiac risk prediction. A study of off-pump coronary artery bypass patients demonstrated that combined models integrating clinical characteristics, radiomics signatures, and non-enhanced CT features of EAT enhanced the accuracy of predicting postoperative atrial fibrillation (POAF)[28]. In aortic stenosis patients, CT radiomics of EAT provided an AUC of 0.80 for discrimination between patients who developed POAF and those maintaining sinus rhythm, with EAT texture showing significantly greater heterogeneity and higher maximum grey-level values in affected patients[29].

Perivascular adipose tissue: Oikonomou et al[30] developed a fat radiomic profile that improved major adverse cardiovascular event prediction with a delta C-statistic of 0.126 (P < 0.001). Lin et al[31] achieved AUC of 0.87 using a pericoronary adipose tissue radiomic phenotype to distinguish acute myocardial infarction from stable coronary artery disease, outperforming attenuation-based models (AUC 0.77).

The perioperative implications of cardiac radiomics are substantial. The 2024 American College of Cardiology/American Heart Association guidelines on perioperative cardiovascular management now include consideration of coronary CT angiography as an alternative to pharmacologic stress testing for preoperative cardiac evaluation[32]. Radiomic enhancement of routine preoperative cardiac CT could simultaneously provide plaque vulnerability assessment, myocardial tissue characterisation, EAT quantification, and cardiovascular risk stratification-all from a single imaging acquisition (Table 4).

Table 4 Cardiac radiomics applications relevant to perioperative risk assessment.

Application	Imaging modality	Key radiomic features	Diagnostic performance	Perioperative relevance
Coronary plaque vulnerability	CCTA	Texture features (GLCM, GLRLM)	AUC 0.73 vs 0.65 (visual)	Identification of high-risk plaques; perioperative MI risk
Acute MI prediction	Cardiac CT	Kurtosis, short-run high gray-level emphasis	AUC 0.90	Perioperative cardiac event prediction
Myocardial pathology classification	Cardiac CT	Multiple texture features	Sensitivity 86%, specificity 81%	Subclinical cardiomyopathy detection
Cardiac amyloidosis detection	Cardiac CT	Radiomic signature	AUC 0.91 (external validation)	Restrictive physiology identification; fluid management
POAF prediction (CABG)	Non-contrast CT (EAT)	EAT texture and shape	Improved combined model	Postoperative arrhythmia risk; anticoagulation planning
POAF prediction (aortic stenosis)	CT (EAT)	Maximum grey-level, heterogeneity	AUC 0.80	Rate/rhythm control prophylaxis
MACE prediction	CCTA (PVAT)	Fat radiomic profile	ΔC-statistic 0.126	Long-term cardiovascular risk stratification
Acute MI vs stable CAD	CCTA (PCAT)	Pericoronary fat radiomics	AUC 0.87 vs 0.77 (attenuation)	Coronary inflammation assessment

CCTA: Coronary computed tomography angiography; AUC: Area under the curve; MI: Myocardial infarction; EAT: Epicardial adipose tissue; POAF: Postoperative atrial fibrillation; CABG: Coronary artery bypass grafting; PVAT: Perivascular adipose tissue; PCAT: Pericoronary adipose tissue; MACE: Major adverse cardiovascular events; CAD: Coronary artery disease.

Pulmonary radiomics: Texture analysis and respiratory reserve

Postoperative pulmonary complications (PPCs) remain a leading cause of perioperative morbidity and mortality. Conventional preoperative pulmonary assessment relies on spirometry, clinical risk scores, and subjective radiographic interpretation. CT radiomics offers an objective, quantitative alternative for characterising pulmonary parenchymal integrity.

Lung texture features-encompassing gray level co-occurrence matrix, gray level run length matrix, and gray level size zone matrix metrics-have been shown to capture information about emphysema severity and pulmonary function beyond what density measurements alone provide[33]. Vegas Sánchez-Ferrero et al[34] showed that simultaneously correcting CT lung-density measurements for volume, noise, and inter-scanner bias improves the quantification of emphysema progression and reliably distinguishes progressors from non-progressors among smokers in a large longitudinal cohort. A 2025 study demonstrated that full-expiratory CT texture-based radiomics improves model performance when combined with conventional gas-trapping measurements for chronic obstructive pulmonary disease (COPD) classification and lung function decline prediction[35].

In the context of preoperative assessment, radiomics features of the target lesion and the lobe-specific CT emphysema score are predictive of pneumothorax occurrence and chest tube insertion requirements after CT-guided lung biopsy[36]. A combination of parenchymal texture features and lung/airway shape on inspiratory CT accurately detected COPD[37], enabling quantitative risk stratification that could complement traditional pulmonary function testing.

Multi-organ radiomics models incorporating erector spinae and whole-lung features have been developed to predict postoperative survival in non-small cell lung cancer patients[38], demonstrating the added value of integrating muscle quality assessment with pulmonary parenchymal analysis-both directly relevant to anaesthetic risk estimation.

Renal radiomics: Non-invasive functional assessment

Preoperative renal function assessment is critical for anaesthetic drug selection, fluid management, and contrast media decision-making. CT radiomics applied to kidney imaging provides non-invasive functional assessment that can complement traditional serum biomarkers.

AI-based automatic estimation of single-kidney glomerular filtration rate using deep learning segmentation (UNETR architecture) and radiomic features from non-contrast CT renal parenchyma has been developed, offering a contrast-free approach that avoids nephrotoxicity risk[39]. This is particularly valuable for perioperative settings where patients may have pre-existing renal impairment or require repeated imaging.

Furthermore, non-enhanced CT-based radiomics combined with kidney volume measurements has shown promise for evaluating chronic kidney disease severity[40], while automated CT measurement of total kidney volume predicts renal function decline following therapeutic interventions, with a 10% or greater decrease in kidney volume at six months predicting 30% or greater estimated glomerular filtration rate decline at 12 months[41].

VASCULAR MORPHOLOGY AND CARDIOVASCULAR RISK STRATIFICATION

Vascular calcification is a recognised independent predictor of perioperative cardiovascular events. CT-based radiomics enables quantitative characterisation of calcification burden and morphology beyond simple calcium scoring.

Automated abdominal aortic calcification (AAC) scores obtained from semi-automated and automated CT analysis have been shown to predict future cardiovascular events with superior accuracy compared to the Framingham Risk Score[12]. Deep learning segmentation combined with CT radiomics enables characterisation of microarchitectural changes within cardiovascular calcification deposits, utilising intensity, size, shape, and texture features[42]. Texture analysis of periaortic adipose tissue reveals distinct features associated with focal aortic calcification, potentially serving as future biomarkers for predicting calcification development[43].

In the carotid vasculature, radiomics has been applied to plaque characterisation for identifying patients at risk of perioperative stroke. A systematic review on radiomics quality in carotid plaque analysis demonstrated growing evidence for differentiating symptomatic from asymptomatic plaques, informing surgical decision-making between carotid endarterectomy and conservative management[44].

CT angiography radiomics of coronary plaques enables detection of rapid plaque progression, with validated radiomic signatures identifying plaques likely to progress, thereby flagging patients who may require perioperative coronary optimisation[45]. For the anaesthesiologist, comprehensive vascular radiomics from preoperative CT can simultaneously inform decisions regarding arterial line placement, blood pressure targets, and the intensity of perioperative cardiovascular monitoring.

INTEGRATING RADIOMIC FEATURES WITH CLINICAL AND LABORATORY DATA

Multi-modal predictive models

The greatest value of radiomics for perioperative risk assessment lies not in imaging features alone but in their integration with clinical, laboratory, and genomic data to create comprehensive predictive models. Recent evidence consistently demonstrates that combined radiomic-clinical models outperform either data source in isolation.

A 2025 study of pancreatic fistula prediction after pancreaticoduodenectomy achieved AUC of 0.93 by integrating preoperative CT radiomics with clinical features, using SHapley Additive exPlanations (SHAP) visualization to transform the model into a clinically interpretable tool[46]. In gastric cancer, light gradient boosting machine (LightGBM)-an ensemble machine learning algorithm-was demonstrated as an effective predictive model for postoperative complications by integrating radiomics features with clinical variables[47]. For gastrointestinal stromal tumour patients, a model integrating clinical, radiomics, and deep learning features accurately predicted postoperative recurrence and metastasis, enabling individualised follow-up strategies[48].

The concept of “surgomics”-the comprehensive integration of preoperative imaging radiomics with intraoperative surgical process data-has been proposed for personalised prediction of morbidity, mortality, and long-term surgical outcomes[49]. Surgomics should be understood as an early conceptual proposal: It extends the radiomics paradigm across the intraoperative phase, but to date it has limited prospective validation, and its incremental value over established perioperative risk models remains unproven. This paradigm nonetheless aligns naturally with the anaesthesiologist’s role as the perioperative physician, coordinating risk assessment, intraoperative management, and postoperative care.

Machine learning architectures for perioperative prediction

Radiomic-clinical integration models draw on machine learning architectures ranging from interpretable classical algorithms to complex deep learning networks. Logistic regression, random forest, and gradient-boosting frameworks such as LightGBM and XGBoost are widely used for structured data combining imaging, clinical, and laboratory variables. Table 5 summarises the principal machine learning approaches used in radiomic-clinical integration models for perioperative outcome prediction.

Table 5 Machine learning algorithms for radiomics-based perioperative risk prediction.

Algorithm	Mechanism	Strengths	Perioperative application examples	Typical AUC range
Logistic regression	Linear decision boundary for binary classification	Interpretable; established statistical framework	Surgical site infection prediction; transfusion requirement	0.70-0.82
Random forest	Ensemble of multiple decision trees	Handles high-dimensional data; resistant to overfitting	Sarcopenia detection; postoperative complication prediction	0.80-0.90
Support vector machine	Optimal hyperplane separation in feature space	Effective in high-dimensional spaces; robust with small samples	Myocardial pathology classification; plaque vulnerability	0.78-0.90
LightGBM/XGBoost	Gradient boosting ensemble methods	High accuracy; fast training; handles missing data	Postoperative gastric cancer complications; pancreatic fistula	0.84-0.93
Deep neural networks	Multi-layer non-linear feature learning	Automatic feature extraction; captures complex patterns	Airway difficulty prediction; cardiac event risk	0.80-0.95
Convolutional neural networks	Spatial feature learning from image data	Direct image input; no manual feature engineering	Airway segmentation; organ volumetry; body composition	0.85-0.96

Area under the curve ranges are indicative values compiled across heterogeneous studies and do not represent single validated estimates. AUC: Area under the curve; LightGBM: Light gradient boosting machine; XGBoost: Extreme Gradient Boosting.

Towards a radiomics-enhanced preoperative assessment framework

Figure 2 illustrates a conceptual framework for integrating radiomics into the preoperative anaesthetic workflow. In this model, routine preoperative CT images undergo automated radiomic feature extraction across multiple anatomical compartments (skeletal muscle, adipose tissue, liver, heart, lungs, kidneys, vasculature). These features are combined with clinical data (age, comorbidities, surgical procedure, ASA status), laboratory values (haemoglobin, creatinine, liver function tests, coagulation profile), and optionally pharmacogenomic data to generate patient-specific risk profiles through validated machine learning models.

Figure 2 Multi-organ radiomics integration framework for precision anaesthesia. CT: Computed tomography; ASA-PS: American Society of Anesthesiologists Physical Status; SMI: Skeletal Muscle Index; SMD: Skeletal muscle density; VAT: Visceral adipose tissue; SAT: Subcutaneous adipose tissue; Vd: Volume of distribution; EAT: Epicardial adipose tissue; PVAT: Perivascular adipose tissue; POAF: Postoperative atrial fibrillation; COPD: Chronic obstructive pulmonary disease; PPCs: Postoperative pulmonary complications; AUC: Area under the curve; NMB: Neuromuscular blocker; HDU: High-dependency unit; ICU: Intensive care unit; SHAP: SHapley Additive exPlanations; LIME: Local Interpretable model-agnostic explanations; GLCM/GLRLM/GLSZM: Texture feature matrices; LightGBM: Light gradient boosting machine; XGBoost: Extreme Gradient Boosting.

Figure 2 multi-organ radiomics integration framework for precision anaesthesia. Conceptual model for transforming a single preoperative CT into a comprehensive patient-specific perioperative risk profile. This framework is a conceptual proposal that requires prospective multicentre validation before clinical use.

The intended output would include: (1) Individualised risk scores for specific complications (cardiac events, PPCs, acute kidney injury, prolonged ileus); (2) Candidate, body composition-informed inputs to anaesthetic drug dosing-currently hypothesis-generating and pending prospective pharmacokinetic validation rather than established dosing recommendations; (3) Airway management strategy guidance; and (4) Postoperative care-level recommendations (ward vs high-dependency vs intensive care).

Workflow responsibility and clinical integration

A recurring barrier to translating the framework in Figure 2 into practice is operational: Who performs the analysis, and at what point in the preoperative pathway? Two deployment models are plausible. In a radiology-led model, automated, IBSI-compliant feature extraction runs as a background process when a preoperative CT is reported; the radiologist validates segmentation quality, and a structured, standardised “radiomics report”-analogous to a coronary calcium-score report- is pushed into the electronic health record (EHR) alongside the conventional narrative report. The anaesthesiologist then consumes a finished, quality-assured product. In an anaesthesiology-facing model, validated, regulatory-cleared point-of-care AI plug-ins run on the picture archiving and communication system (PACS) or the anaesthetic information management system, allowing the anaesthesiologist to generate risk metrics directly during the preoperative consultation.

Each model has trade-offs. The radiology-led model better controls segmentation quality and protocol standardisation but depends on radiology workflow capacity and reporting time; the point-of-care model offers immediacy but transfers quality-control responsibility to clinicians who may lack radiomics training. In both models the timing is favourable, because preoperative imaging is typically acquired days to weeks before surgery, leaving an adequate window for automated analysis; the principal constraints are therefore infrastructure, quality assurance, and clearly assigned accountability rather than processing latency. A pragmatic near-term path is the radiology-led structured report for elective surgery, reserving anaesthesiologist-facing point-of-care tools for later adoption once they are prospectively validated and regulatory-approved.

EMERGING APPLICATIONS: AIRWAY ASSESSMENT AND FRAILTY QUANTIFICATION

AI-based airway assessment

Difficult airway management remains a leading cause of anaesthesia-related morbidity and mortality. Traditional clinical tests (Mallampati score, thyromental distance, mouth opening) demonstrate limited sensitivity (20%-62%) despite reasonable specificity (82%-97%)[50]. AI and imaging-based approaches represent a transformative advance.

A comprehensive review published in Anesthesia and Analgesia (2025) documented that machine learning models using imaging data achieve sensitivity of approximately 80%-90% and specificity of approximately 90%-100%, outperforming classical assessment methods[50]. It must be emphasised that these figures are ranges pooled across methodologically heterogeneous studies, with differing definitions of the difficult airway, differing reference standards, and differing patient populations; they do not represent a single, prospectively validated performance estimate, and should not be interpreted as clinically confirmed metrics. Specific approaches include:

CT-based airway analysis: Grimes et al[51] found a significant positive correlation between a radiographically assessed difficult nostril on preoperative CT and the actual difficulty of nasal intubation, with a positive predictive value of 71.4%, leading the authors to recommend routine review of the CT before planned nasal intubation. Automated CT airway segmentation using convolutional neural networks can eliminate operator bias and substantially reduce segmentation time[50].

Facial image analysis: Hayasaka et al[52] developed a deep learning convolutional neural network model that classified intubation difficulty from facial images with an AUC of 0.864, accuracy of 80.5%, sensitivity of 81.8%, and specificity of 83.3%. Cuendet et al[53] achieved an AUC of 0.81 using automatic facial morphological trait analysis in a database of 970 patients.

CT radiomics for endotracheal tube sizing: A 2025 study developed a predictive model for double-lumen endotracheal tube sizes based on radiomics and AI, achieving accuracy of 0.77, offering a rapid method for airway assessment[54].

Clinical parameter-based machine learning models: Zhou et al[55] applied multiple machine learning and deep learning algorithms to thyroid surgery patients; the gradient boosting model performed best, achieving an AUC > 0.8, accuracy > 90%, and precision of 100% for difficult airway prediction, with age, sex, weight, height, and BMI identified as the most influential predictors.

These approaches suggest that radiomics-enhanced airway assessment-potentially extractable from routine preoperative CT scans of the head, neck, and chest-could provide objective, quantitative airway evaluation to complement bedside clinical tests (Table 6).

Table 6 Comparison of traditional and artificial intelligence/radiomics-based airway assessment approaches.

Assessment method	Sensitivity (%)	Specificity (%)	AUC	Advantages	Limitations
Mallampati score	20-62	82-97	0.55-0.65	Simple; bedside; no equipment needed	Low sensitivity; subjective; operator-dependent
Thyromental distance	25-50	80-95	0.60-0.70	Quick measurement; objective distance	Poor positive predictive value; single parameter
CT-based airway analysis	71-91	85-95	0.80-0.91	Objective 3D assessment; quantitative measurements	Requires CT; radiation exposure; not routine
Facial image deep learning	80-82	84-90	0.81-0.86	Non-invasive; rapid; smartphone-compatible	Requires validation across ethnicities; lighting dependency
Combined ML models	80-90	90-100	> 0.80	Integrates multiple parameters; outperforms individual tests	Limited emergency applicability; dataset-dependent

Reported performance ranges are pooled across heterogeneous studies and do not represent single prospectively validated estimates. CT: Computed tomography; AUC: Area under the curve; ML: Machine learning; 3D: Three-dimensional.

Frailty assessment through imaging biomarkers

Frailty-a multifactorial syndrome of age-related decline in physiologic reserve-is a critical determinant of surgical outcomes in elderly patients but is resource-intensive to assess through comprehensive geriatric assessment[56]. CT-based imaging offers objective frailty surrogates.

A 2025 study of elderly patients undergoing major abdominal oncological surgery found that frailty, diagnosed by comprehensive geriatric assessment, was present in 50.63% of the cohort, with colorectal cancer the most common indication[56]. In that study, AAC and psoas muscle radiodensity were significantly higher in frail patients, whereas other individual radiological parameters (sarcopenia, osteoporosis, renal atrophy) did not differ significantly, and a composite radiological score showed only modest discrimination for frailty[56]. Findings across studies are not uniform: In a separate cohort undergoing emergency laparotomy, osteopenia was the most useful radiological marker for perioperative mortality risk, while sarcopenia, aortic calcification, and kidney volume did not predict poor outcomes[57]. These mixed results indicate that CT body composition analysis can contribute objective frailty surrogates from routinely obtained preoperative imaging, but that no single radiological marker is yet sufficient on its own. This is particularly valuable for anaesthetic planning, as frail patients require modified anaesthetic techniques, lower drug doses, and intensive postoperative monitoring.

CHALLENGES, STANDARDISATION, AND CRITICAL APPRAISAL

The reproducibility challenge

Despite the promising evidence reviewed above, several critical challenges must be addressed before radiomics can be reliably integrated into perioperative practice. The foremost challenge is the “reproducibility crisis” that has afflicted radiomics research[58].

Radiomic features are sensitive to variations in scanner manufacturer, acquisition parameters (kVp, mAs, slice thickness), reconstruction algorithms, and contrast enhancement protocols[8]. A study evaluating feature extraction reproducibility across IBSI-compliant radiomics platforms using a digital phantom found high consistency among common features but identified the necessity for further standardisation in computational algorithms and mathematical definitions due to platform-specific variations[59].

For perioperative applications specifically, this sensitivity is consequential and warrants critical appraisal. Preoperative CT examinations for non-cardiac surgery are acquired for diverse clinical indications and are therefore highly non-uniform in scan timing relative to surgery, contrast phase (non-contrast, arterial, or portal-venous), tube voltage and current, field of view, slice thickness, and reconstruction kernel. Each of these parameters measurably perturbs first-order and, to an even greater degree, higher-order texture features. A radiomic signature trained on one institution’s protocol may therefore degrade substantially when applied to the heterogeneous, opportunistic imaging that characterises real-world surgical populations. Robust perioperative deployment will require harmonisation methods (e.g., ComBat), demonstrably protocol-robust feature subsets, or model retraining and recalibration at each site-none of which is yet standard practice. This is a central reason why research-grade performance figures cannot be assumed to transfer to routine clinical use.

The IBSI

The IBSI, an independent international collaboration established in 2016, has made substantial progress in addressing these challenges[10]. The initiative provides standardised nomenclature, mathematical definitions for over 170 radiomic features, digital phantoms with benchmark values, reporting guidelines, and a subsequently standardised set of reproducible convolutional image filters[60] (Table 7).

Table 7 Image Biomarker Standardisation Initiative standardization framework: Key phases and achievements.

Phase	Timeline	Key achievements	Impact on reproducibility
IBSI phase 1	2016-2020	Standardized definitions for > 170 features; digital phantoms; benchmark values across multiple software platforms	Established common mathematical language; enabled cross-platform comparison
IBSI phase 2	Completed February 2024	Formalized preprocessing workflow (discretization, interpolation, filtering); expanded test phantoms for CT, MRI, PET pipelines	Entropy coefficient of variation decreased from 34% to 7%
Reporting guidelines	Ongoing	Transparent documentation of discretization, filtering, ROI handling, feature selection rationale	Improved study reproducibility and meta-analytic comparability
Compliance tools	Available	Online validation portal; excel-based reference sheets; benchmark feature values	Software platforms (PyRadiomics, LIFEx, SERA, RaCaT) can verify compliance

IBSI: Image Biomarker Standardisation Initiative; CT: Computed tomography; MRI: Magnetic resonance imaging; PET: Positron emission tomography; ROI: Region of interest.

The quantitative impact of IBSI standardisation is substantial: Entropy coefficient of variation decreased from 34% to 7% following IBSI-compliant preprocessing, demonstrating tighter cross-software agreement[10]. Importantly, IBSI complements existing standards including QIBA (acquisition and reconstruction profiles) and DICOM (data exchange and metadata), creating a comprehensive standardisation ecosystem[10].

Cost-effectiveness, infrastructure, and implementation barriers

Radiomics analysis is not a cost-free overlay on existing imaging. It requires specialised software (e.g., PyRadiomics, LIFEx), adequate computational and storage infrastructure, integration middleware, and-critically-trained personnel for segmentation quality control, pipeline maintenance, and model governance. In most hospitals, particularly outside specialised academic centres and in lower-resource health systems, this combination of technology and expertise is not readily available, and no dedicated reimbursement pathway currently exists. The experience of other operator-dependent diagnostic technologies is instructive: Zuo and Ji[61], discussing high-resolution microendoscopy for oesophageal squamous cell neoplasia screening, argued that an expertise gap in the trained-operator base can undermine the theoretical cost-effectiveness of an otherwise promising screening technology. The same caution applies directly to radiomics, whose real-world performance depends on segmentation expertise, protocol control, and sustained data-science support; without deliberate investment in personnel and training, radiomics risks delivering benefit only at well-resourced centres and thereby widening, rather than narrowing, disparities in perioperative care. A formal health-economic evaluation-weighing software, infrastructure, and personnel costs against measurable reductions in complications and length of stay-is a necessary precondition for adoption and has not yet been performed.

Algorithmic fairness, data-source bias, and accountability

Radiomic-clinical models inherit the characteristics of their training data. If a model is developed on a dataset that over-represents or under-represents particular ethnic groups, ages, sexes, or body-habitus extremes, it may perform unevenly across the population to which it is later applied, systematically misclassifying under-represented groups-a particular concern given the ethnicity-dependence of body-composition norms discussed earlier. Site- and scanner-specific data-source bias can further cause a model to learn institutional imaging signatures rather than genuine biology. Equally unresolved is the question of accountability: When a model-informed risk estimate contributes to an anaesthetic decision that precedes an adverse outcome, responsibility must be transparently apportioned among the treating clinician, the institution deploying the tool, and the developer. Mitigating these risks requires demographically representative and multi-institutional training data, prospective subgroup performance reporting, external validation, and explicit governance frameworks established before clinical deployment.

Applicability to special populations

Anaesthetic practice spans populations in which current radiomic evidence is essentially absent. In paediatric patients, body composition, organ proportions, and tissue characteristics change continuously with growth and development, so adult-derived thresholds and signatures cannot be assumed valid. In obstetric patients, pregnancy produces marked physiological and volumetric changes, and the appropriate preference to avoid ionising radiation limits the availability of CT for feature extraction. In emergency surgery, the time required for image transfer, segmentation, and feature extraction may be incompatible with the urgency of the clinical decision. Radiomic models trained on adult elective surgical cohorts should therefore not be applied to these groups without dedicated, population-specific development and validation.

Additional challenges for perioperative translation

Limited perioperative-specific evidence: The overwhelming majority of radiomics research has been conducted in oncologic and cardiovascular contexts, with few prospective studies validated in the perioperative setting of elective or emergency non-cardiac surgery. Most of the predictive-performance figures cited in this review are therefore extrapolated from non-perioperative cohorts, and the apparent superiority of radiomic-clinical models over conventional risk scores has not been confirmed in adequately powered, prospective perioperative trials. This evidence gap is the single most important limitation of the field and should temper the interpretation of every performance estimate presented here.

Associations vs decision-grade utility: Much of the evidence reviewed demonstrates statistical association-for example, between CT-defined sarcopenia or visceral adiposity and adverse postoperative outcomes-rather than proven causal, decision-grade predictive utility. Demonstrating that a radiomic biomarker correlates with risk is not equivalent to demonstrating that acting upon it (by changing anaesthetic technique or drug dose) improves outcome. The latter requires interventional or, at minimum, prospective predictive validation, which is at present almost entirely absent.

Temporal relevance of imaging: Preoperative CT may be obtained weeks to months before surgery. Body composition, hepatic steatosis, and vascular calcification may change in the interim, particularly in patients undergoing neoadjuvant therapy or nutritional optimisation.

Integration with clinical workflows: Real-time radiomic analysis demands automated pipelines integrated with hospital PACS and EHR, requiring significant infrastructure investment.

Regulatory considerations: Radiomic-based clinical decision support tools must undergo regulatory approval pathways, including demonstration of clinical utility and safety in prospective studies.

Interpretability and explainable AI: Many high-performing radiomic models operate as “black boxes”[62]. Clinician reluctance to trust opaque models for high-stakes perioperative decisions is one of the principal barriers to adoption. Explainable AI (XAI) methods-most prominently SHAP and Local Interpretable Model-agnostic Explanations-are therefore not optional refinements but prerequisites for clinical translation. By quantifying how each radiomic and clinical feature contributes to an individual patient’s predicted risk, these methods allow the anaesthesiologist to interrogate the basis of a high-risk designation, to detect when a model is relying on spurious or protocol-related features, and to integrate the prediction with clinical judgement rather than defer to it. The successful use of SHAP to render a pancreatic-fistula prediction model clinically interpretable[46] illustrates the approach, and routine XAI reporting should be expected of any radiomics tool proposed for perioperative use.

AI IN PERIOPERATIVE MEDICINE: THE BROADER CONTEXT

The application of radiomics to anaesthetic planning exists within the broader context of AI adoption in perioperative medicine. A strengths-weaknesses-opportunities-threats analysis published in Frontiers in Digital Health (2024) identified anaesthesiology as a “data-rich medical specialty” uniquely suited for AI implementation[62].

Strengths include AI’s capacity for comprehensive data synthesis, enabling awareness of potential adverse events and personalised physiological assessment. Real-time prediction algorithms can detect changes in vital signs-including intraoperative hypotension and persistent desaturations-that may indicate acute complications[62].

Opportunities include large language models for administrative tasks and unstructured EHR data extraction (with demonstrated 40% time reduction and 18% quality improvement), and real-time closed-loop anaesthetic control algorithms[62].

Weaknesses include the “black box” nature of many AI models, inherited healthcare disparities from training datasets, and high implementation costs[62]. Current AI systems predominantly address singular, linear tasks without comprehensive clinical context-a limitation that multi-modal radiomic-clinical models could help address.

The convergence of precision anaesthesia-incorporating pharmacogenomics, real-time physiologic monitoring, and now quantitative imaging biomarkers-represents a fundamental shift from population-based to individualised perioperative care[9]. Molecular mechanisms driving precision perioperative medicine, including inflammation, metabolism, and neuroimmunomodulation pathways, may eventually be linked to radiomic phenotypes, creating truly integrated multi-omics perioperative risk profiles[63].

FUTURE DIRECTIONS AND RESEARCH PRIORITIES

The translation of radiomics from an oncologic research tool to a clinical perioperative assessment instrument requires focused investigation across several domains.

Prospective perioperative validation studies

Large, multicentre prospective trials are needed to validate radiomics-based perioperative risk models against established clinical risk scores (ASA-PS, Revised Cardiac Risk Index, Surgical Outcome Risk Tool) in diverse surgical populations, including emergency and paediatric settings. This is the single highest research priority for the field, because the present evidence base is overwhelmingly derived from oncological and cardiovascular cohorts rather than genuine perioperative populations.

Pharmacokinetic-radiomic correlations

Studies directly correlating CT-derived body composition parameters with measured drug pharmacokinetics (propofol, remifentanil, neuromuscular blocking agents) are critically needed to establish evidence-based dosing adjustments.

Automated extraction pipelines

Development of fully automated, IBSI-compliant radiomics pipelines integrated with hospital PACS and anaesthetic information management systems, enabling real-time preoperative risk scoring without requiring radiologist intervention.

Multi-compartment analysis from single acquisitions

Algorithms that simultaneously extract body composition, organ function, vascular, and airway features from a single preoperative CT examination, maximising the information yield from existing imaging.

Longitudinal studies

Investigation of temporal stability of radiomic features and the optimal imaging-to-surgery interval for perioperative risk prediction.

Equity and generalisability

Validation across diverse populations, body habitus, and imaging protocols to ensure that radiomic risk models do not perpetuate or amplify existing healthcare disparities, including the establishment of ethnicity-specific body-composition thresholds and prospective subgroup performance reporting.

Health-economic and regulatory pathways

Formal cost-effectiveness evaluation and engagement with regulatory agencies to establish approval frameworks for radiomics-based perioperative clinical decision support tools.

CONCLUSION

Radiomics represents a paradigm-shifting opportunity for preoperative anaesthetic assessment by transforming routine diagnostic imaging into comprehensive physiologic profiles that objectively quantify body composition, organ function, vascular integrity, and airway anatomy. In predominantly oncological and cardiovascular cohorts, combined radiomic-clinical models have reported diagnostic performance (AUC 0.84-0.95) exceeding that of conventional clinical risk scores for selected perioperative complications, and pooled data across heterogeneous studies suggest that AI-based airway assessment can outperform traditional bedside tests; these figures, however, derive from methodologically diverse studies rather than single validated estimates, and prospective perioperative confirmation is largely lacking. The IBSI framework has substantially improved feature-extraction reproducibility, reducing cross-platform variability to clinically more acceptable levels. Significant challenges nonetheless persist, including a nascent perioperative-specific evidence base, sensitivity of radiomic features to heterogeneous CT protocols, unaddressed cost and infrastructure barriers, risks of algorithmic bias, and the absence of validation in paediatric, obstetric, and emergency populations. Radiomics should therefore presently be regarded as a promising, hypothesis-generating direction rather than a clinically ready tool. As imaging technology, AI algorithms, and standardisation frameworks mature-and contingent on rigorous prospective multicentre validation-radiomic biomarker integration into anaesthetic planning holds the potential to transform preoperative evaluation into an objective, individualised, and data-driven precision practice.