Machine Learning-Based Prediction of Clinical Outcomes in Patients With Cancer Receiving Systemic Treatment Using...

Area of Science:

• Oncology
• Digital Health
• Machine Learning

Background:

• Systemic anticancer treatments can lead to clinical adverse events.
• Monitoring patient physical activity offers a potential avenue for early detection of these events.
• Smartphone-based activity tracking provides a scalable and accessible method for continuous patient monitoring.

Purpose of the Study:

• To determine if changes in daily step count, measured via smartphones, can predict clinical adverse events within the following week.
• To evaluate the efficacy of machine learning models in predicting these events using physical activity data.
• To assess the potential for proactive toxicity management strategies in cancer care.

Main Methods:

• A prospective observational cohort study involving cancer patients undergoing systemic treatment.
• Continuous monitoring of daily step count using participants' own smartphones for 90 days.
• Development and validation of machine learning models (elastic net, random forest, neural network) using data from the preceding two weeks to predict adverse events in the next seven days.
• Performance evaluation using Area Under the Curve (AUC).

Main Results:

• The study analyzed 76 patients; 14% experienced unplanned hospitalizations.
• Machine learning models accurately predicted unplanned hospitalizations within 7 days (RF AUC=0.88, NN AUC=0.84, EN AUC=0.83).
• Models demonstrated poor predictive performance for treatment modifications (AUC=0.28-0.51) and overall clinically relevant adverse events (AUC=0.32-0.50).

Conclusions:

• A decline in daily step count is a significant early predictor of upcoming hospitalizations in cancer patients.
• Smartphone-based step count monitoring can facilitate proactive and preventive toxicity management.
• This approach holds promise for improving patient outcomes and reducing healthcare utilization during cancer treatment.