This new approach focuses on patients with non-small cell lung cancer (NSCLC) undergoing systemic therapy. By layering traditional medical records with real-time data such as sleep patterns and heart rate from wearable devices, the model delivers more accurate risk predictions than clinical data alone.
Importantly, it’s built using Bayesian Networks, a machine learning method that emphasizes transparency, enabling clinicians to see not just the risk score, but the underlying patterns and interactions driving it.
Rethinking How We Monitor Cancer Patients
Patients undergoing treatment for advanced NSCLC often face severe side effects that can lead to unplanned urgent care visits or hospitalizations. These events not only disrupt treatment but also negatively impact quality of life and long-term outcomes.
Historically, clinicians have relied on routine checkups and static clinical indicators like lab results or demographics to assess patient risk. But these snapshots can miss subtle signs of decline that occur between visits.
That’s where patient-generated health data (PGHD) comes in. With tools like wearable devices and symptom-tracking surveys, there’s now a way to monitor patients continuously. The real challenge has been figuring out how to meaningfully integrate all this dynamic information into everyday care.
A Layered, Explainable Approach to Risk Prediction
To tackle this, Moffitt researchers designed a study that combined three key data streams from 58 patients during the first 60 days of systemic therapy:
- Standard clinical and demographic data from electronic health records
- Patient-reported outcomes (PROs) collected through the PROMIS-57 quality-of-life survey
- Wearable sensor data (WSD) from Fitbit devices tracking sleep, step count, and heart rate
Together, these sources offered a multidimensional view of each patient's condition, capturing both subjective symptoms and objective behavioral trends.
Rather than using a “black box” AI model, the team turned to Bayesian Networks. This method allowed them to visually map out how different factors like fatigue, physical activity, and clinical variables interact and influence a patient’s risk of needing urgent care. The result was a model that doesn’t just predict risk, but shows why that risk exists.
The difference in performance was clear. A model using only clinical and demographic data performed moderately well. But when PROs and wearable data were added, the predictive accuracy jumped significantly, achieving an Area Under the Curve (AUC) of 0.86. This suggests that it’s not just what data is collected, but how diverse types of data work together that truly matters.
Turning Insight Into Action in Clinical Practice
For clinicians, the potential is significant. Instead of waiting for symptoms to escalate into emergencies, this model could flag at-risk patients early, allowing for timely, targeted interventions that are less invasive and more effective.
And because the model is explainable, it doesn’t just produce a number; it offers context. That’s key for clinical adoption. Providers can understand how a patient's activity level, reported symptoms, and clinical metrics are contributing to their risk, which supports better decision-making.
While the results are promising, the researchers acknowledge that this is an early-stage study. Conducted at a single center with a modest sample size, it lays the groundwork for larger, multi-institutional validation efforts. The team also plans to expand the model by incorporating additional data layers, such as molecular markers, to further refine its accuracy and adaptability.
A Step Toward More Proactive Cancer Care
Ultimately, this study demonstrates how integrating clinical data with real-time, patient-driven insights can improve not just prediction, but care itself. The model offers a clear, interpretable way to identify patients at risk, bridging the gap between episodic clinical visits and the continuous reality of living with cancer.
This approach pushes cancer care toward something smarter, more responsive, and more personalized by giving clinicians a fuller picture of their patients' day-to-day health and doing so in a way that's both accurate and transparent.
Journal Reference
Gonzalez et al., 2025. Using Bayesian Networks to Predict Urgent Care Visits in Patients Receiving Systemic Therapy for Non-Small Cell Lung Cancer. PubMed, 9, e2400315–e2400315. DOI:10.1200/cci-24-00315. https://ascopubs.org/doi/10.1200/CCI-24-00315
Disclaimer: The views expressed here are those of the author expressed in their private capacity and do not necessarily represent the views of AZoM.com Limited T/A AZoNetwork the owner and operator of this website. This disclaimer forms part of the Terms and conditions of use of this website.