The clinical challenge. Lung cancer remains the leading cause of cancer-related death worldwide, claiming approximately 1.8 million lives in 2022. For patients without driver mutations or immune checkpoint expression, platinum-based chemotherapy remains the primary treatment, yet predicting who will survive remains highly difficult.
Limitations of current staging. Although TNM staging is a recognized prognostic factor, it explains less than 30% of differences in patient outcomes. Genomic biomarkers can refine prognosis, but expensive genetic testing limits their clinical application, creating an unmet need for accessible, multidimensional mortality prediction tools.
The biopsychosocial angle. This study incorporated patient-reported symptom burden alongside clinical data, recognizing that how patients feel during chemotherapy - captured through validated questionnaires - contains prognostic information often missed by laboratory tests alone.
Study goal. The team developed and validated a machine learning model combining demographic characteristics, environmental exposures, blood biomarkers, clinical features, and longitudinal patient-reported outcomes to predict 1-, 3-, and 5-year all-cause mortality in 1,278 lung cancer patients receiving chemotherapy.
Large retrospective cohort. The study enrolled 1,278 patients with histologically confirmed lung cancer who received platinum-based chemotherapy at Guangzhou Chest Hospital between 2017 and 2019. Patients were followed biannually by telephone for up to 11 years, with survival status and date of death recorded for all participants.
Multisource data integration. The model drew on five categories of predictors: demographic characteristics (age, sex, BMI), environmental exposures (smoking history, passive smoke), clinical information (TNM stage, histology), immunohistochemical markers, and prechemotherapy blood biomarkers including white blood cell count, hemoglobin, D-dimer, and tumor markers such as CEA, CA12-5, CA19-9, CYFRA21-1, and C-reactive protein.
Patient-reported outcomes. Patients completed the MD Anderson Symptom Inventory for Lung Cancer (MDASI-LC) at admission and across the first four chemotherapy cycles. This validated 22-item questionnaire captures symptom severity on a 0-10 scale. The Karnofsky Performance Status and Zubrod Performance Score were also collected to assess functional status.
Longitudinal trajectory modeling. Rather than using a single symptom score, the team applied group-based trajectory modeling (GBTM) to characterize how symptom burden evolved over time. This analysis identified three distinct MDASI-LC trajectory groups (high, medium, and low burden), capturing the dynamic nature of chemotherapy side effects as a prognostic variable.
Comprehensive algorithm testing. Five machine learning approaches were tested: random survival forest (RSF), gradient boosting machine (GBM), survival support vector machine (survival-SVM), elastic net (Enet), and stepwise Cox regression. Each was trained with multiple hyperparameter configurations, yielding 13 distinct base models.
Exhaustive pairwise combinations. To explore ensemble strategies, all possible two-model combinations of the 13 base models were generated, creating a total of 84 predictive models. The concordance index (C-index) calculated on the held-out test set was used to identify the single best-performing model.
Winning combination. The optimal model was a combination of forward stepwise Cox regression and gradient boosting machine (GBM), achieving a C-index of 0.702 (95% CI 0.652-0.753) on the test set. Bootstrap internal validation across 1,000 resamplings confirmed robustness with an average C-index of 0.682.
Feature attribution. SHAP (SHapley Additive exPlanations) analysis and restricted cubic splines (RCS) were applied to interpret how each of the 21 retained features contributed to mortality predictions, including identifying critical numerical thresholds beyond which risk accelerated nonlinearly.
Improving accuracy over time. On the test set, the model achieved AUCs of 0.740, 0.777, and 0.915 for 1-, 3-, and 5-year mortality predictions respectively. The substantial improvement at 5 years likely reflects both the model's strength in capturing long-term risk factors and the smaller number of surviving patients at that time point.
Calibration quality. Calibration curves showed strong agreement between predicted mortality probabilities and observed outcomes. Brier scores (which measure prediction error) were 0.20, 0.18, and 0.11 for 1-, 3-, and 5-year predictions on the test set, reflecting progressively better calibration at longer time horizons.
Clinical utility confirmed. Decision curve analysis demonstrated positive net benefit for the model when risk thresholds were between 0.40-0.69, 0.62-0.99, and 0.72-0.99 for 1-, 3-, and 5-year mortality predictions, meaning the model adds value over treating everyone or no one across a wide range of clinical decision thresholds.
Survival stratification. Using the model's risk scores to divide patients into high-risk and low-risk groups produced significantly different Kaplan-Meier survival curves (p less than 0.01), confirming the model could meaningfully separate patients by prognosis across all three time horizons.
Demographic risk factors. Age emerged as a critical predictor with a threshold at 60 years - below which mortality risk was relatively low - accelerating substantially beyond that point. Male sex, current or former smoking history, and passive smoke exposure were all independently linked to higher mortality, underscoring the importance of lifestyle factors in chemotherapy outcomes.
BMI and the obesity paradox. BMI showed an inverse relationship with mortality risk, with a critical threshold at 20.7 kg/m2. Below this value, mortality risk was markedly higher, consistent with the "obesity paradox" in cancer patients - where mild overweight provides survival advantage possibly through protective adipokines and metabolic reserves during treatment.
Tumor and staging factors. T3/T4 stage, advanced nodal involvement (N2-N4), and distant metastatic stage (M2) all significantly increased mortality risk. CD56 positivity and small cell lung cancer histology were associated with shorter survival times, while no significant difference emerged between adenocarcinoma and squamous cell carcinoma subtypes.
Blood biomarker thresholds. Prechemotherapy levels of CYFRA21-1, CA12-5, CA19-9, white blood cell count, D-dimer, and CEA all showed nonlinear relationships with mortality, with specific inflection points beyond which risk accelerated sharply. Hemoglobin below 120 g/L markedly increased risk, while C-reactive protein showed a linear risk increase with an inflection at 15.3 mg/L.
Symptom burden predicts survival. Patients in the high-symptom burden trajectory group (median MDASI-LC score 45.4) had significantly higher mortality risk compared to those in the low-burden group (median score 8.2). This finding confirms that how patients feel during chemotherapy - their day-to-day symptom experience - carries genuine prognostic information beyond clinical measurements.
Why this matters clinically. Integrating the MDASI-LC into routine clinical follow-up creates an opportunity to identify high-risk patients early through regular symptom monitoring. Patients with escalating symptom burden could trigger earlier clinical review, supportive care interventions, or treatment modifications before mortality risk fully materializes.
CD56 and immunohistochemistry. CD56 positivity - a marker of neuroendocrine differentiation - was associated with increased mortality risk in this model, though its prognostic significance remains debated in the broader literature. This finding highlights neuroendocrine features as worthy of consideration in individualized prognostic assessment for lung cancer.
Actionable implications. The model's integration of patient-reported outcomes alongside clinical data suggests that oncologists should routinely collect and track symptom burden trajectories during chemotherapy. Combined with prechemotherapy blood biomarkers and staging, this multimodal approach could enable more personalized supportive care and treatment planning.
A practical mortality prediction tool. The study successfully developed an interpretable machine learning model combining 21 multisource features to predict all-cause mortality at 1, 3, and 5 years in chemotherapy-treated lung cancer patients. The model demonstrated acceptable discrimination, calibration, and clinical utility across all time horizons.
Clinical value. By enabling personalized risk stratification before and during treatment, the model could help oncologists identify high-risk patients who might benefit from intensified monitoring, earlier supportive interventions, or consideration of alternative treatment regimens to improve outcomes.
Important limitations. The retrospective, single-center design limits external validity and generalizability. Using the test set to select the best model from 84 candidates may inflate performance estimates. The stability of identified risk thresholds under different model configurations has not been validated, and prospective multicenter replication is needed.
Future directions. External validation in diverse patient populations, prospective data collection incorporating serial symptom assessments, and model refinement using larger, multicenter datasets are necessary next steps before clinical deployment of this tool in routine oncology practice.