Artificial Intelligence Remote Patient Monitoring for Predicting Overall Survival for Patients Undergoing Radical...

Previous studies have highlighted the benefits of using artificial intelligence-powered remote patient monitoring (AI RPM) in detecting health changes across various disease cohorts. However, the use of AI RPM for identifying health deteriorations in patients following major surgical procedures remains underexplored.

This exploratory analysis of a prospective trial aims to assess how AI RPM can enhance the predictive performance of 35-month post-radical cystectomy (RC) mortality risk. Our approach highlights the importance of RPM features in improving prediction accuracy and provides interpretable model outputs to enhance clinical understanding and transparency.

We used patient data from a multicenter RC trial conducted in the United Kingdom for model training and validation. Two gradient-boosted machine learning models were developed: one using only clinical-pathological (CP) features and another incorporating both CP and remote patient monitoring (RPM) features (CP+RPM). RPM features are measured by wrist-worn pedometers and surveys. The predictive accuracy of the CP+RPM model was compared with both the CP model and a clinically used nomogram, both of which relied solely on traditional clinical features. We used 200 bootstrap iterations, with 70% of the data used for training and 30% for testing. Shapley Additive Explanations were applied to interpret model results and provide insights into the relative importance of features, improving transparency and understanding of the predictions.

A total of 252 patients (33 deaths) from 9 UK centers were included in the analysis. We examined 108 RPM features and 24 CP features for model training. In correlation analysis, only 9 CP features showed coefficients larger than 0.1, compared with 36 RPM features with stronger correlations. The CP+RPM model achieved an area under the receiver operating characteristic curve of 0.77, reflecting a 9% and 10% absolute (13% and 15% relative) improvement over the CP and nomogram models, respectively. Similarly, it outperformed in terms of the area under the precision-recall curve, with a score of 0.44, marking a 6% and 17% absolute (16% and 63% relative) increase compared with the CP and nomogram model. Shapley Additive Explanations analysis revealed that the most significant contributors to mortality prediction were mobility-related RPM features, such as the 30-second chair-to-stand test results and daily step count variance, which reflected the general activity levels of individuals.

Our study demonstrates that RPM features significantly enhance long-term survival prediction for post-RC patients, offering a valuable addition to traditional clinical data. The integration of AI with RPM enables more individualized and dynamic tracking of recovery, improving prediction accuracy and fostering a patient-centered care model that has the potential to be applied across a broader range of surgeries and conditions.