TL;DR
This study explores the use of smartphone and wearable sensor data to predict daily symptom burden in patients undergoing pancreatic surgery. By applying machine learning, researchers achieved promising accuracy in forecasting high symptom days, as well as specific symptoms like pain, fatigue, and diarrhea, suggesting a future for passive digital monitoring in cancer care.
What Problem Does This Paper Address?
Cancer treatments, including surgery and chemotherapy, often lead to a range of debilitating symptoms that significantly impact patients' quality of life. A major challenge in oncology is that clinicians frequently miss these symptoms, leading to delayed or inadequate management. This can result in prolonged suffering and functional impairment for patients. While patient-reported outcomes (PROs) are valuable, long-term tracking can be burdensome for patients, leading to decreased compliance over time. The paper addresses this by investigating whether passively collected data from everyday devices like smartphones and wearables can provide an objective and less burdensome way to monitor symptom burden, especially during the critical perioperative period.
Methods
The study employed a prospective longitudinal design, collecting data from 44 patients scheduled for pancreatic surgery. Participants used an Android smartphone with the AWARE app for two weeks before surgery, throughout their inpatient recovery, and for up to 60 days post-discharge. The AWARE app passively collected sensor data (e.g., movement, location, device usage) and also prompted daily self-reported symptom ratings on a scale of 0-10 for 10 different symptoms. Additionally, participants wore a Fitbit Charge 2 to collect data on activity, sleep, and heart rate.
Key to their approach was the extraction of day-level behavioral features from both smartphone and Fitbit data. These features, totaling 213, encompassed mobility, activity patterns, sleep, screen time, heart rate, and communication. Machine learning, specifically the Light Gradient Boosting Machine (LightGBM) algorithm, was used to build predictive models. These models were trained to classify the next day as having a high or low symptom burden, adjusted for each individual's baseline symptom levels. The study also focused on predicting specific symptoms: pain, fatigue, and diarrhea.
Datasets
The study utilized data collected from 44 participants undergoing pancreatic surgery. This dataset comprised:
- Smartphone sensor data: Collected via the AWARE app, including movement, approximate location, device use metadata (calls, SMS), ambient light, and noise levels.
- Wearable sensor data: Collected via a Fitbit Charge 2, including activity, sleep patterns, and heart rate.
- Patient-reported symptom data: Daily ratings of 10 physical and psychological symptoms.
The paper does not introduce a new benchmark dataset but rather uses the collected data for training and evaluating their predictive models.
Key Insights and Ideas
- Passive sensing as a viable alternative to burdensome PROs: The study strongly supports the idea that passively collected sensor data can offer a less intrusive way to monitor patient well-being compared to frequent self-reporting, which often suffers from compliance issues.
- Behavioral patterns as digital biomarkers for symptom burden: The research highlights that specific patterns in physical activity, sleep, and heart rate, captured by everyday devices, can serve as "digital biomarkers" indicative of a patient's symptom burden.
- Perioperative period is crucial for monitoring: The study emphasizes the importance of monitoring during the perioperative phase, a time when complications are common and early detection can prevent readmissions.
- Individualized symptom thresholds are key: Adjusting predictions based on each patient's typical symptom level (residual scores) is crucial for accurate classification, acknowledging the inherent variability in symptom experience.
Key Results
- Overall Symptom Burden Prediction: The LightGBM model achieved 73.5% accuracy in predicting whether the next day would be a high symptom day, outperforming baseline models.
- Specific Symptom Prediction:
- Diarrhea: 79.0% accuracy
- Fatigue: 75.8% accuracy
- Pain: 79.6% accuracy
- Important Features: For overall symptom burden, the most influential features included recent symptom scores, days since surgery, average symptom burden, and various activity/heart rate metrics. For specific symptoms, location data was important for diarrhea, while step and sleep features were key for fatigue and pain.
Limitations and Future Work
- Small Sample Size: The authors acknowledge that the model performance, while promising, may not be adequate for immediate clinical implementation due to the relatively small sample size (n=44).
- Study-Provided Devices: The use of study-provided smartphones might differ from patients using their own devices, potentially impacting real-world generalizability.
- Impact of Surgery and Hospitalization: The significant physiological and behavioral changes associated with major surgery and hospitalization could influence model performance.
- Future Directions:
- Larger Sample Sizes: Future studies should involve larger cohorts to improve model robustness.
- Personal Devices: Utilizing patients' own smartphones and wearables would enhance real-world applicability.
- Subgroup Analysis: Developing models for specific subgroups of patients (based on demographics, clinical factors) could yield better results than a single population model.
- Real-time Implementation: The ultimate goal is to use these models for real-time prediction to trigger timely symptom management interventions and personalized support.
- Longer Monitoring Periods: Extending monitoring beyond the current perioperative window to cover longer treatment and survivorship phases is also suggested.
Summary updated