Federated learning inverts the usual logic of data science. Instead of shipping raw data to a central server, each participating site trains a model on its own local data and transmits only the resulting model parameter updates, such as learned network weights, to an aggregator that combines them into a shared global model. In the study, two data nodes, one hosting the Lifelines cohort and the other the Rotterdam Study, ran local training independently on the Vantage6 federated learning platform, while a separate aggregator node computed the combined weights using Federated Averaging, or FedAvg, a method that weights each site’s contribution by its number of training samples. The exchanged weights were encrypted end-to-end between the participating organizations so that even the central server could not read them.

The two cohorts involved were strikingly different, and that heterogeneity is precisely what makes the result noteworthy. Lifelines is a large population-based study of healthy aging in the northern Netherlands, contributing 148,230 eligible participants with cardiovascular outcomes derived from self-reported questionnaires. The Rotterdam Study, begun in 1990 in Rotterdam’s Ommoord district, contributed a much smaller cohort of 10,155 participants, but its outcomes were identified through continuous digital linkage with hospital discharge records and general practitioner files, providing exact event dates and complete follow-up. In Lifelines, event dates for heart attack, stroke, or heart failure had to be approximated as the midpoint between the last event-free assessment and the assessment at which the event was first reported, introducing imprecision in both the event indicator and the timing.

Before any learning could begin, the researchers had to solve a more mundane but formidable problem: making the two datasets speak the same language. They used the MyDigiTwin data harmonization pipeline, which converts raw tabular cohort data into patient-centered Common Data Format files, applies dataset-specific mapping rules to produce HL7 FHIR-compliant resources following the ZIB profile used in Dutch personal health environments, and then flattens those resources into the tabular structure the federated algorithm requires. Crucially, each cohort was harmonized independently at its own site, and no raw data were transferred between sites at any stage. The final predictor set comprised 12 variables consistently available in both cohorts: sex, age, systolic and diastolic blood pressure, HDL, LDL, and total cholesterol, estimated glomerular filtration rate, creatinine, smoking status, prevalent type 2 diabetes, and prevalent hypertension.

The prediction model itself was DeepSurv, a deep learning extension of the classical Cox proportional hazards model that replaces the linear risk score with a feed-forward neural network, allowing non-linear associations between risk factors and time-to-event outcomes to be learned from data. The network was deliberately compact, with an input layer of 12 neurons, two hidden layers of 16 neurons each, ReLU activations, batch normalization, and dropout of 0.20, amounting to roughly 500 trainable parameters. Each aggregation round therefore involved exchanging a parameter vector of only that size. Local training used the Adam optimizer with a mini-batch size of 1,024 and a learning rate of 1 × 10⁻⁴, with 100 local epochs per update iteration across 20 update iterations, for a total of 2,000 local epochs per node in the federated setting.

Performance was measured with the C-statistic, a measure of how well a model discriminates between individuals who do and do not experience events, reported as the mean and 95% confidence interval across 10 independent runs. In the Rotterdam Study, the primary evaluation cohort because of its complete follow-up, the C-statistic rose from 0.723 (95% CI: 0.714–0.732) for the locally trained model to 0.737 (95% CI: 0.727–0.747) with federated training. In Lifelines, the increase was more modest, from 0.791 (95% CI: 0.783–0.799) to 0.795 (95% CI: 0.785–0.804). The improvement was larger in the Rotterdam Study, plausibly because its smaller sample size and lower baseline performance allowed it to benefit more from the information contributed by the vastly larger partner cohort.

The researchers were careful to rule out a tempting alternative explanation: that the federated model simply won by training longer. When the local Rotterdam Study model was trained for 2,000 epochs, matching the total local training budget of the federated setting, its mean C-statistic was 0.724 (95% CI: 0.709–0.739), essentially unchanged from the 100-epoch result. The gain therefore came from the other cohort’s data rather than from additional computation. The convergence trajectories also differed between the sites: the Rotterdam Study captured most of its improvement at the very first aggregation and then plateaued between 0.733 and 0.738, while Lifelines improved gradually and near-monotonically across all 20 iterations, consistent with the aggregated model being weighted heavily toward the Lifelines node.

Perhaps the most intriguing finding concerns data quality. A central worry in federated healthcare AI is that a large cohort with noisy outcomes could drag down a smaller one with precise data. Here, the opposite appeared to happen: despite Lifelines’ self-reported outcomes and interval-based event timing, its inclusion through federated learning was associated with higher performance on the Rotterdam Study, suggesting that the sheer volume of information in the larger cohort compensated, at least partially, for the uncertainty in its labels. The authors caution that unreported events in Lifelines would enter the partial likelihood as censored observations, potentially weakening the estimated associations, and that because FedAvg weights by sample size, any such weakened signal could propagate into the aggregated model in proportion to Lifelines’ dominant contribution. The observed improvement may therefore be a conservative estimate with respect to this source of uncertainty.

The study also comes with clearly stated boundaries. Privacy preservation here means keeping participant-level data within local institutional environments under the cohorts’ governance arrangements, including a Joint Controller Agreement and a Data Protection Impact Assessment, not a formal guarantee against inference attacks on the exchanged model parameters; techniques such as secure aggregation and differential privacy could further reduce those risks, though both may add computational overhead and, in differential privacy’s case, reduce performance. Only two nodes participated, both in the Netherlands, so generalizability to more institutions or to populations with different demographic and ethnic composition remains untested. Non-cardiovascular deaths were treated as censoring events rather than competing risks, which may overestimate absolute risk in the older Rotterdam Study participants, and the Cox partial likelihood’s reliance on risk sets means the federated optimization is not equivalent to fitting a single model on pooled data.

The practical implications are nonetheless striking. The entire federated training schedule ran in roughly 30 minutes on a single CPU core for a Lifelines-sized dataset and about 2 minutes for a Rotterdam Study-sized one, on standard virtual machines without dedicated accelerators, indicating that the real cost of this kind of collaboration lies in orchestrating tasks and governance rather than computation. Whether a discrimination gain of this magnitude would change clinical decisions around statin or blood pressure treatment thresholds would require reclassification analysis, decision-analytic measures such as net benefit, and external validation. But as a demonstration that heterogeneous, real-world cohorts can jointly train deep survival models under existing data protection frameworks, the study offers a template for the multi-institutional, privacy-preserving medical AI that regulators and patients alike have been waiting for, with future directions pointing toward imaging-derived biomarkers, longitudinal predictors, and aggregation methods that account for differences in cohort size and data quality.

Subject of Research: Federated deep learning for privacy-preserving cardiovascular disease risk prediction using the Lifelines and Rotterdam Study cohorts

Article Title: Federated deep learning for privacy-preserving cardiovascular disease risk prediction

Article References: Mo, H., Smits, D., Birhanu, M. A., Leening, M. J. G., Bos, D., van der Harst, P., & Bron, E. E. (2026). Federated deep learning for privacy-preserving cardiovascular disease risk prediction. Frontiers in Applied Mathematics and Statistics, 12, Article 1923866. https://doi.org/10.3389/fams.2026.1923866

Image Credits: AI Generated

DOI: 10.3389/fams.2026.1923866

Keywords: federated learning, cardiovascular disease, risk prediction, deep learning, DeepSurv, Lifelines, Rotterdam Study, privacy, FedAvg, C-statistic, Vantage6, survival analysis

The research team, led by investigators at Tianjin Union Medical Center and Nankai University in China, assembled a substantial dataset to train and test their algorithms. Drawing on the Surveillance, Epidemiology, and End Results database, the large American cancer registry, they identified 3,511 patients who had undergone surgery for gastrointestinal stromal tumors. That cohort was randomly divided at a ratio of seven to three into a training set, used to teach the models, and an internal validation set, used to check their performance on data they had never seen. To guard against the possibility that the models had merely learned quirks of the American registry, the team added a second, entirely independent test: 219 patients treated at their own medical center in Tianjin, whose outcomes served as an external validation cohort.

Before any modeling began, the researchers confronted a problem that plagues every predictive study in oncology: which of the many variables recorded for each patient actually matter? They applied three separate feature selection methods to sift through the candidate predictors, allowing the data itself to determine which variables earned a place in the final models. Notably, the variables that survived this screening process differed between the two prediction tasks. The factors that best forecast death from any cause were not identical to those that best forecast death specifically from the tumor, a distinction that has real clinical consequences, because many patients with these tumors ultimately die of something other than their cancer.

With the variables chosen, the team built four distinct survival models, each representing a different branch of modern machine learning. The random survival forest, an ensemble method that grows many decision trees adapted to censored survival data, competed against a gradient boosting machine, which builds predictions by sequentially correcting the errors of simple models; CoxBoost, a boosting algorithm built directly on the classical Cox proportional hazards framework; and DeepSurv, a neural network approach that extends the Cox model with deep learning. Each model was evaluated with a battery of quantitative metrics, including Harrell’s concordance index, which measures how well a model ranks patients by risk, along with time-dependent areas under the receiver operating characteristic curve, precision-recall analysis, average precision, and decision curve analysis, which estimates the net clinical benefit of acting on a model’s predictions.

The verdict was clear. The random survival forest delivered the most stable predictive performance across both endpoints and both validation settings. In the internal validation cohort, its concordance index reached 0.735 for overall survival and 0.776 for cancer-specific survival, figures that indicate meaningful discrimination between high- and low-risk patients. The model also produced favorable time-dependent AUC values, meaning its accuracy held up across the follow-up period rather than degrading after the first year or two. When the models were confronted with the external cohort of Chinese patients, the random survival forest again demonstrated reliable discrimination and calibration, the latter meaning that when the model predicted, say, a seventy percent probability of surviving five years, roughly seventy percent of such patients did in fact survive that long. Calibration is often the Achilles heel of machine learning models in medicine, which can rank patients correctly while assigning them wildly inaccurate probabilities, so the external confirmation on a demographically distinct population carries particular weight.

Perhaps the most consequential part of the study is not the raw accuracy but the interpretability. Black-box algorithms have faced justified skepticism in clinical medicine, where a prediction that cannot be explained is a prediction that cannot be trusted, audited or acted upon. The researchers addressed this by applying SHAP, or SHapley Additive exPlanations, a technique borrowed from cooperative game theory that assigns each input variable a quantified contribution to every individual prediction. Rather than asking only whether the model is accurate, SHAP reveals why it made each call, showing which features pushed a given patient’s estimated survival up or down and by how much. This transparency allowed the team to identify the dominant drivers of prognosis and to verify that the model was learning biologically plausible relationships rather than statistical artifacts.

The SHAP analysis produced a striking asymmetry between the two survival endpoints. For overall survival, age emerged as the predominant factor, which makes intuitive sense: a model predicting death from any cause will naturally weight the passage of time and its physiological consequences heavily. For cancer-specific survival, however, the picture shifted decisively toward tumor biology, with tumor size and mitotic rate, the measure of how rapidly cancer cells are dividing under the microscope, dominating the predictions. This divergence is more than a statistical curiosity. It tells clinicians that the question a patient is really asking matters. If the concern is how long the patient will live overall, age and its associated comorbidities loom large. If the concern is whether the tumor itself will prove fatal, the aggressiveness of the tumor, captured by its size and its proliferative activity, is what counts. A single one-size-fits-all risk score would blur this distinction.

Translating the model from a research finding into a usable clinical instrument, the team built an online survival calculator based on the random survival forest. The tool allows a clinician to enter an individual patient’s characteristics and receive personalized, time-specific estimates of overall and cancer-specific survival, replacing the coarse risk categories of traditional staging systems with continuous, individualized probabilities. For a fifty-eight-year-old patient with a six-centimeter tumor and a moderate mitotic count, the calculator can generate a survival curve specific to that profile, giving both surgeon and patient a far more granular picture of the road ahead than any nomogram or risk table currently in routine use. Such tools also have the potential to refine follow-up schedules, concentrating surveillance on patients the model flags as high risk while sparing low-risk patients unnecessary imaging and anxiety.

The study arrives amid a broader movement to bring explainable artificial intelligence into oncology, and its design reflects lessons learned from earlier efforts that stumbled. By combining multiple feature selection methods, testing four competing algorithms, validating externally on a separate population, and interrogating the winning model with SHAP, the researchers assembled a methodological chain of evidence that goes beyond a single impressive accuracy figure. The work was approved by the ethics committee of Tianjin Union Medical Center, and the authors declared no competing interests. The study was supported by the Research Project of the Tianjin Municipal Education Commission.

Limitations remain, as they do in any registry-based study. The SEER database, for all its scale, records a defined set of variables and may not capture every factor that influences outcomes, such as the use of targeted therapies like imatinib, molecular mutations in KIT or PDGFRA, or details of surgical technique. The external validation cohort, while independent, comprised 219 patients from a single center, and broader prospective testing will be needed before such a calculator becomes standard practice. Still, the direction of travel is unmistakable. For a tumor whose behavior has long defied simple classification, a model that can weigh tumor size, mitotic rate, age and other variables together, quantify their contributions for each patient, and deliver an honest, calibrated survival estimate represents a genuine step toward precision follow-up care. The era in which a postoperative gastrointestinal stromal tumor patient’s prognosis could be summarized by a three-tier risk label may be drawing to a close, replaced by algorithms that speak in probabilities, show their work, and fit in a browser window.

Subject of Research: Explainable machine learning models for predicting postoperative survival in gastrointestinal stromal tumor patients

Article Title: Prediction of overall and cancer-specific survival in postoperative patients with gastrointestinal stromal tumors based on explainable models

Article References: Chen, Z., Zhao, J., Wang, N., Zhang, J., Song, P., Tian, W., Luo, N., Zhang, S., & Cao, L. (2026). Prediction of overall and cancer-specific survival in postoperative patients with gastrointestinal stromal tumors based on explainable models. BMC Cancer. https://doi.org/10.1186/s12885-026-17022-x

Image Credits: AI Generated

DOI: 10.1186/s12885-026-17022-x

Keywords: gastrointestinal stromal tumor, machine learning, random survival forest, SHAP, overall survival, cancer-specific survival, SEER database, DeepSurv, prognosis, risk stratification, explainable AI, BMC Cancer