Machine Learning Models Show Complementary Strengths in Predicting Gastric Cancer Survival
A single-institution study reveals that classical Cox models and modern machine learning algorithms each excelled at different performance metrics in forecasting outcomes after stomach cancer surgery, offering an instructive lesson about the tradeoffs between interpretability and complexity.
In a finding that offers an instructive lesson about clinical artificial intelligence, researchers have demonstrated that different predictive models showed complementary strengths in forecasting survival after radical gastrectomy for gastric cancer. The single-institution study, published in BMC Cancer, found that while a classical Cox model achieved the highest concordance index, a Random Survival Forest excelled at time-dependent predictive accuracy, suggesting that model selection involves tradeoffs between different performance metrics and clinical interpretability.
The research addresses a fundamental challenge in gastric cancer care: two patients with seemingly similar tumors can face radically different futures. Accurate prognostic prediction enables clinicians to tailor surveillance strategies and treatment intensity to individual risk profiles, potentially sparing low-risk patients from unnecessary interventions while identifying high-risk patients who may benefit from aggressive therapy.
The research team, led by Jun Wu, Luqing Zhou, Haidi Chang, and Xuhui Liao from The First Affiliated Hospital of Lishui University and Lishui People's Hospital, analyzed 200 patients who underwent radical gastrectomy at their institution. Working in collaboration with Tian Qin and Junxuan Zhu from Guangzhou LBP Medical Technology Co., Ltd., they compared three predictive approaches: the traditional Cox proportional hazards model (CoxPH), Random Survival Forest (RSF), and Gradient Boosting Survival Analysis (GBSA).
Comparative Performance Across Metrics
The Cox proportional hazards model—a statistical technique developed in the 1970s—achieved the highest discrimination with a C-index of 0.733 ± 0.096 in head-to-head comparison with more computationally intensive competitors. However, the Random Survival Forest demonstrated exceptional time-dependent AUC values of 0.911 at 12 months, 0.933 at 36 months, and 0.950 at 60 months, indicating that its ability to distinguish survivors from non-survivors actually improved as follow-up extended.
The cohort consisted of 79.0% male patients with a mean age of 65.4 years. Over a median follow-up period of 64.2 months, 91 of the 200 patients (45.5%) died. The models identified two independent prognostic factors: age (hazard ratio 1.097, 95% CI 1.068-1.145, P<0.001) and N stage lymph node involvement (hazard ratio 1.498, 95% CI 1.100-2.322, P=0.035).
Dramatic Risk Stratification
The models enabled risk stratification that separated patients into groups with dramatically different outcomes. Using the CoxPH model, the low-risk group (n=67) achieved a 91.0% five-year survival rate, while the intermediate-risk group (n=66) reached 57.6%, and the high-risk group (n=67) only 22.4%—a 68.6 percentage point spread between the extremes.
This stratification capability could offer healthcare providers a data-driven framework for individualizing post-surgical surveillance intensity and adjuvant therapy decisions, the researchers suggest. Low-risk patients with 91% five-year survival might be spared aggressive overtreatment, while high-risk patients with 22% survival could be identified for intensified intervention.
The study utilized 5-fold stratified cross-validation for internal validation, incorporating variables including age, sex, TNM stage, Lauren classification, differentiation grade, tumor size, lymphovascular invasion, perineural invasion, and treatment information. The research was approved by the Ethics Committee of Lishui People's Hospital, with informed consent waived due to the retrospective design using de-identified data.
Limitations and Future Directions
The authors declared no competing interests and received no external funding. Importantly, they acknowledge that external validation in diverse populations and independent institutions remains essential before clinical implementation—a critical caveat for any healthcare organization considering deployment. The single-institution design with internal validation only raises concerns about optimism bias in the reported performance metrics, underscoring the need for rigorous testing in independent patient cohorts from other medical centers before these models can be considered ready for clinical use.
The findings validate a reproducible methodology that balances different dimensions of predictive performance with clinical interpretability. The classical model's strong performance alongside the Random Survival Forest's exceptional time-dependent accuracy addresses fundamental considerations in clinical adoption: physicians can understand and explain risk calculations based on interpretable models to patients, potentially facilitating regulatory approval and clinical integration, while ensemble methods may offer advantages for long-term surveillance planning.
The study offers a template for integrating machine learning responsibly in oncology by comparing classical and modern algorithms under rigorous cross-validation and translating model outputs into clinically actionable risk categories. However, the single-institution limitation means that generalizability to other patient populations, healthcare settings, and geographic regions remains unproven until external validation studies are completed.
Sources
Google Search
Follow Health AI Central on Google
Choose Health AI Central as a preferred source, and Google is more likely to show our reporting in Top Stories.
Reporting & accountability
How we maintain accuracy
Health AI Central separates reported facts, analysis, and public-source data. Readers can review our standards, request a correction, or consult the relevant research methodology before relying on a claim.