The clinical dilemma at the heart of the research is well known to oncologists. Patients with locally advanced gastric cancer frequently receive neoadjuvant chemotherapy before surgery to shrink tumors and improve resection outcomes. After the operation, many are offered additional adjuvant chemotherapy, yet the evidence supporting this second round of treatment in patients who have already received preoperative therapy remains contested. Some trials suggest benefit, others show marginal gains, and clinicians are left weighing toxic side effects against uncertain rewards for each individual patient. Current decisions rely heavily on post-surgical pathological staging, a blunt instrument that captures little of the biological and clinical nuance that determines whether a particular patient will relapse.

To address this gap, a multi-institutional team led by researchers at the National Cancer Center/Cancer Hospital of the Chinese Academy of Medical Sciences, together with colleagues from Tianjin Medical University Cancer Institute and Hospital, Beijing Friendship Hospital, and the Cancer Hospital of China Medical University, assembled a retrospective cohort of 1,150 patients treated with neoadjuvant chemotherapy and radical gastrectomy across four Chinese centers. Rather than relying on a single algorithm, the team embraced a large-scale combinatorial strategy. They first employed eleven different machine learning learners, each of which identified its own optimal subset of predictive features from the clinical data. These eleven feature subsets were then crossed with the eleven learners, producing 121 candidate prediction models that competed against one another for predictive supremacy.

The evaluation process was rigorous and multidimensional. The researchers assessed each candidate using the concordance index, a standard measure of how well a survival model ranks patients by risk; time-dependent receiver operating characteristic curves, which capture discrimination at specific follow-up horizons; time-dependent calibration curves, which test whether predicted probabilities match observed outcomes over time; and decision curve analysis, which quantifies the net clinical benefit of acting on the model’s predictions. Out of the 121 contenders, one model clearly rose above the rest: the GAMB-AORSF model, an acronym that combines a Generalized Additive Models via Gradient Boosting-selected feature subset with an Accelerated Oblique Random Survival Forest learner.

The technical architecture of the winning model reflects two complementary strengths. Gradient-boosted generalized additive models are highly effective at screening large sets of candidate variables and selecting a compact, informative feature subset without imposing rigid linear assumptions. The accelerated oblique random survival forest, in turn, is a tree-based ensemble method designed specifically for censored survival data. Unlike conventional random forests, which split data on single variables at each node, oblique random survival forests consider linear combinations of variables, allowing them to capture more complex interaction structures in the data. The “accelerated” designation refers to computational optimizations that make this demanding approach feasible at scale. This pairing proved exceptionally powerful for modeling time-to-recurrence outcomes.

The performance numbers are striking. In the training cohort, the GAMB-AORSF model achieved a concordance index of 0.864, and it maintained robust discrimination in two independent validation cohorts, with C-indices of 0.813 and 0.789. A C-index of 0.5 would correspond to random guessing, while 1.0 represents perfect ranking; values above 0.8 in external validation are rarely achieved in oncology prediction models, particularly those built from routinely collected clinical variables rather than expensive molecular profiling. The model also successfully stratified patients into distinct risk groups whose survival trajectories diverged substantially, providing a practical foundation for treatment personalization.

The most clinically consequential finding, however, came from the model’s use as a treatment-selection instrument. Because decisions about adjuvant chemotherapy are not randomized in routine practice, the researchers applied inverse probability of treatment weighting, a statistical technique that simulates the balance of a randomized trial by reweighting treated and untreated patients according to their probability of receiving treatment. Within the model-defined risk strata, a clear pattern emerged. High-risk patients derived significant survival benefits from adjuvant chemotherapy: across the cohorts, the treatment extended three-year restricted mean survival time by five to seven months and produced an absolute reduction in recurrence risk of 14 to 20 percent. In contrast, low-risk patients showed no significant survival improvement from additional chemotherapy, implying that many of these patients may be enduring weeks of toxic treatment with little to show for it.

Restricted mean survival time deserves particular attention as an outcome measure. Unlike hazard ratios, which can be difficult to interpret when treatment effects vary over time, restricted mean survival time quantifies the average amount of life or disease-free time gained over a fixed horizon, expressed in familiar units of months. For a high-risk patient, gaining five to seven months of cancer-free survival is a clinically meaningful benefit that most would consider worth the side effects of chemotherapy. For a low-risk patient, an unmeasurable benefit against real toxicity argues for de-escalation. The model effectively converts a population-level debate into an individual-level decision.

The study’s use of SHAP values, a game-theoretic approach to explaining machine learning predictions, further addresses a common criticism of artificial intelligence in medicine: the black-box problem. By quantifying each feature’s contribution to individual predictions, SHAP analysis allows clinicians to see why a particular patient was classified as high or low risk, fostering the transparency needed for clinical adoption. The researchers also evaluated their model against the PROBAST framework for prediction model risk of bias and reported their work following TRIPOD guidelines, signaling attention to methodological standards that many published clinical prediction tools lack.

The implications extend beyond gastric cancer. Neoadjuvant chemotherapy followed by surgery is increasingly the standard of care for multiple solid tumors, and in each setting the same question arises: after a response to preoperative therapy, does everyone still need postoperative treatment? The Chinese team’s approach, combining exhaustive model search with causal inference methods to estimate treatment effects within risk strata, offers a template that could be adapted to esophageal, rectal, and other cancers. The framework of using machine learning not merely to predict outcomes but to guide de-escalation decisions represents a shift toward genuinely precision-guided perioperative oncology.

Caveats remain. The study is retrospective, and despite external validation across multiple centers, the findings arise from a Chinese patient population treated largely with regimens such as SOX, FLOT, and DOC, which may limit immediate generalizability to other populations and treatment protocols. Prospective validation, ideally in a randomized trial design where the model is used to stratify treatment assignment, will be needed before the GAMB-AORSF model can change guidelines. Integration with emerging biomarkers such as circulating tumor DNA, which the authors note as a future direction, could further sharpen the model’s risk distinctions.

Funding for the work came from the National Natural Science Foundation of China, the Beijing Natural Science Foundation, the Capital Health Development Research Special Fund, and Tianjin medical research programs. The corresponding authors are Quan Xu, Guoliang Zheng, and Yantao Tian, with Xu Liu, Peng Jin, Peng Wang, Xinxin Shao, and Haikuo Wang contributing equally as first authors. The study received institutional review board approval and written informed consent from all participants, and the authors declared no competing interests.

For now, the message for patients and clinicians is one of cautious optimism. A tool that reliably separates those who need continued treatment from those who do not could spare hundreds of thousands of patients worldwide unnecessary chemotherapy each year, while concentrating intensive treatment on those most likely to relapse. As machine learning models like GAMB-AORSF move toward prospective testing, the era of one-size-fits-all postoperative care in gastric cancer may finally be drawing to a close.

The clinical stakes of the problem are considerable. Patients with sepsis occupy a peculiar intersection of risk factors for thrombosis. The same inflammatory storm that characterizes sepsis activates the coagulation cascade, endothelial damage is common, and prolonged immobility during intensive care compounds the danger. Traditional risk assessment tools, such as the widely used Padua and Caprini scores, were developed largely for general hospitalized populations and, according to the study’s authors, have shown limited effectiveness when applied to sepsis patients, whose physiology differs substantially from that of the average medical or surgical ward patient. This gap in predictive accuracy has real consequences: anticoagulation prophylaxis, the standard preventive treatment for VTE, carries its own risks of bleeding, and clinicians need to know precisely which patients warrant aggressive prevention. A tool that can reliably stratify thrombotic risk in this population could meaningfully shift the balance of preventive care.

To build their models, the research team, led by Chenglong Liang and Chen Zhou of the First Affiliated Hospital of Wenzhou Medical University, along with colleagues at several affiliated institutions, retrospectively collected data from ICU sepsis patients treated at the hospital. The dataset comprised 1,824 patients, of whom 235 developed venous thromboembolism during their ICU stay, an incidence that underscores just how prevalent the complication is in this population. The researchers divided the data temporally rather than randomly, a methodological choice designed to mimic real-world deployment. Patients admitted between 2021 and 2023 were split into a training cohort containing 70 percent of cases and a validation cohort containing the remaining 30 percent, while patients admitted from January to June 2024 served as an entirely independent test cohort, providing a rigorous check on whether the models could generalize to patients admitted after the model-building period.

Before any modeling could begin, the team confronted two of the most common pitfalls in clinical machine learning: irrelevant or redundant predictors and severely imbalanced classes. Because only a minority of sepsis patients develop VTE, a naive model could achieve superficially high accuracy simply by predicting that no one is at risk. To counter this, the researchers applied the Synthetic Minority Over-sampling Technique, or SMOTE, to the training cohort, generating synthetic examples of the minority class, patients who did develop VTE, so that the algorithms could learn the distinguishing features of both groups rather than defaulting to the majority. Feature selection proceeded in stages: univariate logistic regression was first used to screen candidate predictors, after which stepwise regression based on the Akaike Information Criterion and the least absolute shrinkage and selection operator, known as Lasso, refined the variable set. The Lasso method, in particular, shrinks the coefficients of uninformative variables toward zero, effectively performing automated variable pruning while guarding against overfitting.

The result of this selection process was a compact and clinically interpretable set of six predictors: patient age, length of stay in the ICU, duration of hospitalization before ICU admission, use of glucocorticoids, use of analgesic medications, and serum D-dimer levels. Each of these variables makes intuitive clinical sense. Older patients tend to have more compromised vascular systems and reduced mobility. Longer ICU stays mean more time exposed to the prothrombotic conditions of critical illness. Pre-ICU hospitalization duration may reflect disease severity and prolonged immobilization before intensive management began. Glucocorticoids, staples of sepsis management in many protocols, are known to influence coagulation pathways, and analgesic use may serve as a proxy for both disease burden and immobility. D-dimer, a fibrin degradation product routinely measured in critically ill patients, is a direct biochemical marker of active clot formation and breakdown, making it perhaps the most mechanistically direct of all the predictors.

Armed with these six features, the researchers constructed and compared six machine learning models spanning the spectrum of complexity from classical statistics to modern ensemble methods: logistic regression, decision tree, random forest, support vector machine, gradient boosting machine, and extreme gradient boosting. Each model was evaluated with a battery of standard performance metrics, including receiver operating characteristic curves and the area under them (AUC), calibration curves, positive and negative predictive values, true positive and true negative rates, accuracy, F1-score, and the Brier score, which quantifies the overall accuracy of probabilistic predictions. The AUC, a measure of a model’s ability to discriminate between patients who will and will not develop VTE, served as the primary basis for comparison, with values closer to 1.0 indicating stronger discrimination.

XGBoost emerged as the top performer in both the validation and test cohorts, demonstrating that its sophisticated approach, which builds an ensemble of decision trees sequentially, with each new tree trained to correct the errors of the ensemble so far, captured subtle patterns in the sepsis data that simpler models missed. In the validation cohort, the XGBoost model achieved an AUC of 0.837, with a 95 percent confidence interval of 0.79 to 0.883, placing it in the range commonly regarded as good clinical discrimination. Its performance held up reasonably well in the temporally separated test cohort, where the AUC was 0.792 (95 percent CI 0.716 to 0.869). For comparison, the other five models were also tested, but none matched XGBoost’s combination of sensitivity and specificity across the two evaluation sets.

The detailed performance metrics in each cohort reveal both the strengths and the realistic limitations of the tool. In the validation cohort, XGBoost correctly identified 73.3 percent of patients who went on to develop VTE (the true positive rate, or sensitivity) and correctly cleared 75.9 percent of those who did not (the true negative rate, or specificity). In the test cohort, sensitivity rose to 82.4 percent while specificity dipped to 70.2 percent. The negative predictive values are particularly striking: 94.8 percent in the validation cohort and 97.4 percent in the test cohort, meaning that when the model predicts a patient is at low risk, that reassurance is very often correct. Positive predictive values, at 32.4 percent and 22.8 percent respectively, were more modest, a mathematical consequence of the relatively low baseline incidence of VTE, but this matters less in a screening context where the goal is to cast a reasonably wide net while avoiding overwhelming false alarms.

The clinical implications of a high negative predictive value deserve emphasis. In an ICU setting, where every intervention carries cost and risk, a model that can reliably rule out VTE risk in the majority of patients would allow clinicians to concentrate prophylactic attention, whether that means more aggressive anticoagulation, enhanced surveillance imaging, or mechanical measures such as compression devices, on the smaller group flagged as high risk. The authors conclude that the XGBoost model demonstrated superior discriminatory ability compared with the other models and has the potential to assist clinical healthcare professionals in identifying high-risk VTE patients among those with sepsis in the ICU. Because the study was retrospective, the next step for this line of research would naturally be prospective validation, ideally in multi-center settings, to confirm that the model’s performance translates across different hospitals, patient mixes, and clinical practices.

The study also reflects a broader trend in critical care medicine: the migration of machine learning from academic curiosity to bedside decision support. Ensemble tree methods like XGBoost have become workhorses of clinical prediction because they handle nonlinear relationships, interactions between variables, and heterogeneous data types with minimal manual feature engineering. What distinguishes the current work is its disciplined methodology, temporal data splitting to test real-world generalization, principled feature selection through Lasso and information-criterion-based stepwise regression, class balancing with SMOTE, and evaluation across a comprehensive set of metrics rather than a single headline number. The research was supported by the National Natural Science Foundation of China and several Zhejiang provincial programs, and the study was approved by the Ethics Review Committee of the First Affiliated Hospital of Wenzhou Medical University, with informed consent waived due to its retrospective nature. As hospitals worldwide grapple with rising sepsis admissions and the accumulating evidence that VTE significantly worsens outcomes in these patients, tools of this kind may soon become a quiet but consequential part of the intensive care workflow, flagging the vulnerable before the clot forms.