Search

Prostate cancer metastasis to bone and lymph nodes marks a critical transition to incurable disease, with five-year survival dropping dramatically compared to localized disease. Early identification of patients at high risk of metastasis enables timely intensification of treatment, including androgen deprivation therapy, salvage radiation, or systemic therapies. Current deep survival models that integrate serial PSA measurements and genomic risk scores achieve high predictive accuracy for time-to-metastasis but operate as black boxes, providing no explanation for why a particular patient is predicted to have early or late metastasis. Clinicians cannot trust or act upon predictions without understanding which PSA features or genomic markers drive the risk assessment. We present an explainable deep survival framework that combines a deep survival model for time-to-metastasis prediction with Integrated Gradients attribution, a method that distributes the model's hazard prediction among input features. The framework produces patient-specific explanations showing how each serial PSA value and each genomic score component contributes to the predicted metastasis hazard. The framework consists of three core components: (1) a deep survival model (DeepSurv architecture) with a PSA time-series encoder and genomic risk encoder, (2) Integrated Gradients attribution computed over the hazard function, and (3) visualization tools for individual and population-level interpretations. Integrated Gradients attributes the predicted hazard to individual PSA measurements across time and specific genomic markers, enabling clinicians to distinguish between risk driven by rapid PSA kinetics versus high genomic risk scores. This interpretability transforms a black-box survival prediction into an actionable clinical decision support tool.

Chemotherapy remains a cornerstone of cancer treatment, but it is frequently associated with severe toxicities, with 30–80% of patients experiencing grade 3–4 adverse events that may require dose reduction, treatment delays, or hospitalization. While machine learning models have shown strong potential in predicting chemotherapy-related toxicities using electronic health records, genomic data, and clinical variables, most existing approaches generate only point estimates (e.g., a single risk probability) without quantifying uncertainty, limiting their clinical reliability. Such miscalibrated predictions can lead to overconfident risk underestimation or excessive caution, both of which may negatively impact treatment decisions and patient outcomes. This manuscript proposes a conceptual framework that integrates neural network-based toxicity prediction with adaptive conformal prediction to produce calibrated, patient-specific prediction intervals with formal coverage guarantees. The framework combines a feedforward neural network for risk estimation, a non-conformity score to measure how atypical a patient is relative to the training data, and an adaptive calibration mechanism that updates interval thresholds over time to reflect shifts in patient populations and clinical practice. This design enables narrower intervals for well-represented, predictable cases and wider intervals for atypical or high-uncertainty patients, thereby making prediction reliability explicit. Importantly, the method provides finite-sample coverage guarantees without requiring distributional assumptions, ensuring that true toxicity outcomes fall within the predicted intervals at a user-specified confidence level. By transforming point predictions into uncertainty-aware, clinically interpretable intervals, the framework supports more robust, risk-stratified decision-making in chemotherapy planning and moves toward safer, more trustworthy AI-assisted oncology care.

Fall risk in aging populations is a modifiable health concern, with mobility patterns changing over time due to factors like frailty, comorbidities, and medication. Smartwatch accelerometers provide a privacy-sensitive way to monitor gait and movement outside clinical settings. However, federated learning, which supports privacy by keeping sensor data local, faces challenges in aging populations due to concept drift from gradual mobility decline, which can invalidate static models. This article proposes a federated continual learning framework to adaptively maintain fall risk prediction models using smartwatch data. The system includes local models that combine feature extraction with temporal sequence modeling, continual learning to prevent forgetting, and a federated server for privacy-preserving coordination. It aims to support personalized fall risk monitoring, reduce concept drift, and enable scalable deployment in senior care settings, with clinical validation necessary for real-world assessment.