Predicting prostate cancer mortality using UK Biobank data: a multimodal machine learning approach.

Abstract

Prostate cancer remains a major cause of cancer-related mortality among men worldwide; however, it remains challenging to predict disease progression in individuals. While numerous tools attempt to forecast the likelihood of developing the cancer, fewer studies focus specifically on mortality prediction using population-scale clinical datasets. UK Biobank data were used to develop population-scale predictive models of prostate cancer mortality. The study assessed medical records, blood work, physical measurements, and genetic information of approximately 167 000 males aged 50 and above. A feedforward neural network was compared with logistic regression to determine their effectiveness. The neural network demonstrated substantially stronger predictive performance than logistic regression, achieving a receiver operating characteristic-area under the curve of 0.767 compared with 0.633 and a precision-recall-area under the curve of 0.271 compared with 0.153, respectively. The analysis revealed that age and genetics were the primary influencing factors. Decision curve analysis demonstrated minimal net benefit across clinically relevant thresholds. These findings support the utility of multimodal risk modeling while highlighting challenges in clinical translation.