Predict Urgent Care Visits in NSCLC Patients with 80% Accuracy

A study published in JCO Clinical Cancer Informatics demonstrates that machine learning models incorporating patient-reported outcomes and wearable sensor data can predict which patients with non–small cell lung cancer are most at risk of needing urgent care during treatment. The study was led by researchers and clinicians at Moffitt Cancer Center.

Patients undergoing systemic therapy for non-small cell lung cancer often experience treatment-related toxicities that can result in unplanned urgent care visits. In this study, Moffitt researchers tested whether integrating multiple sources of patient-generated health data, including self-reported quality-of-life surveys and wearable device metrics such as sleep and heart rate, could improve predictions beyond standard clinical and demographic information.

The team used explainable machine learning approaches called Bayesian Networks to build predictive models among 58 patients monitored with Fitbit devices and surveyed through a questionnaire. Machine learning models that included patient-reported outcomes and wearable sensor data significantly outperformed models based on clinical data alone on ability to distinguish between high-risk and low-risk patients.

"By combining information patients provide about their symptoms with continuous monitoring from wearable devices, we can better identify who is most at risk for treatment complications," said Brian D. Gonzalez, Ph.D., lead author and a researcher in Moffitt's Department of Health Outcomes and Behavior. "Our goal is to give clinicians tools to intervene earlier, improve patient experiences and potentially prevent hospitalizations."

The findings suggest that integrating multidimensional data into machine learning models may enhance personalized cancer care and allow providers to proactively address toxicities before they escalate. While the study was limited to a single center and a modest sample size, researchers say the approach holds promise for broader application.

"What makes this approach powerful is not only the accuracy of the predictions, but also the ability to understand why the model reaches those predictions," said Yi Luo, Ph.D., co-author and researcher in Moffitt's Department of Machine Learning. "By using explainable machine learning methods, we can see how factors like symptom reports, sleep quality and lab results interact to influence risk. This transparency is critical for building trust with clinicians and ensuring that the models can be used to guide real world decisions in cancer care."