Utilizing Machine Learning to Assess the Role of Physical Function in Predicting Mortality Among Cancer Survivors

Background and/or Objectives: The objective of this study was to develop a machine learning (ML)-based mortality prediction model for cancer survivors using physical function measures.

Design: This study is a retrospective analysis of medical charts.

Setting : This study was conducted at a cancer institute in Southeastern Florida. Physical function was assessed by Patient Reported Outcome Measurement Information Systems (PROMIS)-Physical Function short form and PROMIS-Fatigue short form, Timed Up and Go (TUG) test, the 30-second Sit-to-Stand (STS30) test, the Four-Stage Balance (4SB) test, and grip strength (GS) using a Jamar handheld dynamometer. We used 4 ML algorithms for this study: logistic regression model, random forest, Extreme Gradient Boosting (XGBoost), and neural network. We chose to use these ML algorithms because they have been shown to identify patients at risk of short-term mortality based on structured electronic health record data.

Participants : Participants include cancer survivors referred for an initial cancer rehabilitation consultation with a physiatrist.

Interventions: This study does not have any intervention.

Main Outcome Measures: The primary outcome variable of this study was mortality.

Results: Of the 1832 cancer survivors included in the study, 140 were deceased (mortality rate=7.6%). The XGBoost classifier, identified as the most robust model, predicted patient death with an accuracy of 92%, a recall of 93%, a precision of 91%, an F1-score of 90%, and an area under the curve (AUC) of 0.94.

Conclusions: The XGBoost algorithm had the highest accuracy and precision to predict mortality of cancer survivors using physical function measures. This model has the potential to serve as a clinical decision support system, aiding clinicians to improve health outcomes and lower mortality rates in cancer survivors.