Integrating tumor location into artificial intelligence-based prognostic models in cancer

Abstract

This letter is a commentary on the findings of Huang et al, who emphasize the prognostic value of tumor location in gastric cancer. Analyzing data from 3287 patients using Kaplan-Meier and multivariate Cox models, the authors found that the tumor location correlated with patient prognosis following surgery. Patients with tumors situated nearer to the stomach’s proximal end were associated with shorter survival periods and poorer outcomes. Notably, gender-based differences in tumor markers, particularly carbohydrate antigen 72-4, further highlight the need for sex-specific influence on the tumor location. Despite increasing recognition of tumor location as a prognostic factor, its role remains unclear in clinical prediction models for various cancers. This letter highlights the potential of incorporating tumor location into artificial intelligence -based prognostic tools to enhance prognostic models. It also outlines a stepwise framework for developing these models, from retrospective training to prospective multicenter validation and clinical implementation. In addition, it addresses the technical, ethical, and interoperability challenges critical to successful real-world prognosis.

Key Words: Tumor location; Prognosis; Artificial intelligence; Artificial intelligence-based prognostic tools; Clinical prediction models

Core Tip: This study highlights the prognostic significance of tumor location in gastric cancer, showing that proximal tumors are associated with worse survival outcomes. Gender differences, particularly in carbohydrate antigen 72-4 expression, further influence prognosis. The letter proposes integrating tumor location into artificial intelligence-based clinical prediction models to improve prognostic accuracy. It outlines a stepwise framework for model development, multicenter validation, and clinical implementation, while addressing critical technical, ethical, and interoperability challenges for real-world application.

TO THE EDITOR

The article by Huang et al[1] highlights the clinical significance of tumor location in gastric cancer (GC) outcomes. Methodologically, the study used a retrospective design, analyzing data from a total of 3287 gastric adenocarcinoma cases who underwent gastrectomy from 2008-2019. Kaplan-Meier survival analysis and multivariate Cox regression analyses were utilized to evaluate overall survival and adjusting for covariates. The main finding is that tumor location is an independent prognostic factor for GC. Tumors in the proximal stomach are significantly linked to worse outcomes than distal tumors (log-rank test, P < 0.001). This may be due to delayed diagnosis, more aggressive pathology, and anatomical challenges. Furthermore, clinicopathological indices such as gender, age, serum levels of carcinoembryonic antigen, carbohydrate antigen 19-9 (CA19-9) and CA72-4 exhibited various influences on tumor location. Although statistically robust, the study has limitations. First, the study did not consider the potential influence of Helicobacter pylori infection on tumor location[2]. Similarly, GC cases are significantly affected by geographic region, ethnicity, and dietary habits, all of which can influence tumor localization[3,4]. These important variables were not included in the current study. This letter goes beyond the current finding by underscoring the potential value of integrating tumor location into artificial intelligence (AI)-driven prognostic tools. It outlines a stepwise approach from model development to clinical application, and discusses key technical, ethical, and interoperability considerations necessary for effective real-world clinical implementation.

GENDER DIFFERENCES IN PROGNOSTIC FACTORS AND BROADER ONCOLOGIC IMPLICATIONS

An intriguing discovery in the study is the gender-based variation in prognostic indicators. Huang et al[1] found that CA72-4 could serve as a tumor marker in female patients, particularly for assessing the incidence of GC subsites. However, in male patients, CA72-4 was not a useful prognostic marker for GC subsites. This difference may be attributed to estrogen signaling, hormone receptor expression, and sex-specific environmental exposures. Accumulated evidence has shown that sex differences influence tumor development, immune responses, and treatment efficacy[5-7]. For example, in hepatocellular carcinoma, estrogen is believed to confer a protective effect in females by modulating inflammatory and proliferative signaling pathways[8]. Similarly, in lung cancer, different mutational spectra and treatment responses have been observed between men and women[9-11]. Such disparities emphasize the need for sex-stratified analyses in both prognostic model development and decision-making.

TUMOR LOCATION AS A PROGNOSTIC FACTOR BEYOND GC

In addition to GC, tumor location[12] affects prognosis in many other types of cancers (Figure 1)[13-17]. For example, in colorectal cancer, a well-established survival discrepancy exists between right-sided and left-sided tumors[18]. This could be due to proximal colon tumors often present with microsatellite instability and BRAF mutations, leading to poor prognoses[19]. Similarly, in pancreatic cancer, tumors in the head of the pancreas are generally diagnosed earlier due to bile duct obstruction. However, pancreatic tail tumors are often diagnosed later and worse[20]. Furthermore, in breast cancer, tumor location has been linked to recurrence patterns due to varying lymphatic spread[21]. These studies reinforce the anatomical and biological significance of tumor location across multiple cancer types. Through incorporating tumor location into prognostic models, it may directly guide clinical decision-making and improve patient care.

Figure 1 Tumor types where anatomical location affects prognosis. It illustrates multiple cancer types where tumor prognosis varies with anatomical location. Examples include brain tumors (brainstem vs non-eloquent regions), lung cancer (central vs peripheral), intrahepatic cholangiocarcinoma (perihilar vs peripheral), renal cell carcinoma (hilar vs peripheral), pancreatic cancer (head vs body/tail), esophageal cancer (upper vs lower), breast cancer (upper lateral quadrant vs other quadrants), gastric cancer (cardia vs non-cardia), colorectal cancer (right-sided vs left-sided), and melanoma (head and neck, trunk, upper extremities, lower extremities). GEJ: Gastroesophageal junction.

AI-DRIVEN PROGNOSTIC MODELS FOR CANCER PATIENTS

The advancement of AI-based models has significantly enhanced the development of clinical prognostic tools. Tumor location is not only linked to disease progression and survival outcomes but also reflects underlying molecular, anatomical, and therapeutic differences[22]. Given the multifaceted implications of tumor location, there is a pressing need to incorporate this factor into AI-driven prognostic models. These models can integrate both structured data such as age, sex, laboratory values, histologic subtype, TNM staging and tumor location, and unstructured data including imaging findings, and pathology reports[23]. In addition, advanced algorithmic approaches should be considered when building these models. These approaches may include ensemble learning such as gradient boosting and random forest, as well as deep learning techniques such as convolutional neural networks and recurrent neural networks[24,25]. Moreover, hybrid algorithms combining multiple modalities may be utilized to enhance model performance[26]. In this way, it would better capture the prognostic impact of tumor location across various cancer patients.

AI-BASED MODEL DEVELOPMENT AND VALIDATION STRATEGY

A rigorous, step-by-step strategy is essential for the development, validation, and deployment of AI models in clinical settings. In step 1, the algorithm should be developed using data from a single-center retrospective cohort. Internal validation can be achieved through k-fold cross-validation and bootstrapping[27]. Generally, the performance could be evaluated by area under the receiver operating characteristic curve, sensitivity, specificity, and calibration plots[28]. Step 2 would involve prospective multicenter validation. The model should be applied in diverse healthcare environments, including urban academic centers, community hospitals, and rural clinics. This aims to test its generalizability. By evaluating outcomes across different populations and care settings, external validity can be confirmed[29]. Notably, standardization of data input and processing pipelines is essential to minimize variability across sites, particularly when deploying across different Electronic Health Record (EHR) systems or institutions[30]. Step 3 would focus on clinical integration. This includes embedding the model within EHR systems to support real-time clinical decision-making[31]. Furthermore, a collaboration with clinical guideline committees to evaluate the feasibility of AI-derived risk scores is urgently needed. In addition, transparent documentation of the model’s parameters, decision logic, and model updates will facilitate clinician trust and accountability (Figure 2)[32]. For example, Choi et al[33] developed a deep learning-based model using multi-classification DNN to predict overall survival in 1020 patients with laryngeal squamous cell carcinoma. The model incorporated multiple clinical variables, including tumor location (glottis, supraglottis, subglottis), and demonstrated improved performance over traditional TNM-based models (C-index = 0.859 vs 0.504). This highlights the prognostic relevance of tumor location and supports its inclusion in AI-based prediction models.

Figure 2 Standardized pipeline for development, validation, and clinical deployment of artificial intelligence models in clinical settings. The schematic illustrates a three-step framework: (1) Model development involves training using single-center retrospective data, with internal validation via k-fold cross-validation and bootstrapping, and performance assessed by area under the receiver operating characteristic curve, sensitivity, specificity, and calibration plots; (2) Prospective multicenter validation applies the preliminary model across heterogeneous clinical settings to evaluate generalizability and establish external validity. This is crucial to ensure the model performs equally good across external populations; and (3) Clinical integration incorporates the validated model into Electronic Health Record systems, with guideline committee collaboration and transparent documentation of parameters, decision logic, and updates to enable real-time clinical use and support provider confidence. AUROC: Area under the receiver operating characteristic.

TECHNICAL CONSIDERATIONS AI-BASED MODELS AND IMPLEMENTATION CHALLENGES

Several key challenges must be addressed to ensure successful clinical translation. First, interoperability is crucial as AI systems must be compatible with existing hospital information technology infrastructure and data standards[34]. Second, ethical and regulatory frameworks must be established appropriately. For example, Patient privacy must be protected in accordance with applicable national laws. Informed consent must be completed for AI-driven decision making[35]. Furthermore, models should be explainable to gain clinician trust and facilitate adoption. Explainable AI methods, like Shapley Additive exPlanations values, which allocate importance values to model features, help make complex model decisions more transparent and interpretable to end users[36].

Another technical challenge in AI-based clinical applications is data imbalance, especially in rare tumor sites, which may lead to biased predictions and poor generalization. Recent studies have shown that ensemble learning methods, such as Balanced Random Forest or Easy Ensemble, combined with advanced resampling techniques like SMOTEENN (a hybrid technique combining SMOTE and Edited Nearest Neighbors), can effectively mitigate class imbalance and enhance model robustness across diverse cancer datasets[37]. Incorporating these approaches during model development is crucial for improving the reliability of predictions, particularly for underrepresented patient groups.

CONCLUSION

To further enhance cross-institutional model development, federated learning has emerged as a promising approach enabling collaborative AI training without sharing raw patient data. This is especially relevant in global oncology, where privacy laws and institutional barriers often hinder data sharing. For example, Pati et al[38] demonstrated that federated learning frameworks can match the performance of centrally trained models in multi-institutional cancer imaging tasks, while ensuring patient data remains within each institution’s local environment. Although anatomical location offers valuable clinical context, relying on it alone may miss important biological differences. Tumors at the same location can behave differently depending on their molecular subtypes or tumor microenvironment. Combining these features may lead to more accurate and comprehensive AI models. In summary, the manuscript by Huang et al[1] significantly advances our understanding of the prognostic relevance of tumor location in GC. This study can guide personalized care and highlights a broader pattern across other malignancies, which further emphasizes the relevance of tumor location in cancer prognosis. Integrating tumor location into AI-based predictive models represents a logical and clinically impactful strategy. Through a well-established model development framework, robust validation, and careful attention to ethical and technical considerations, AI-based prognostic scores have the potential to revolutionize prognosis and improve patient outcomes in cancer patients.