Toxicities drive unplanned care. Patients receiving systemic therapy (chemotherapy, immunotherapy, or targeted therapy) for non-small cell lung cancer (NSCLC) frequently experience side effects severe enough to require unplanned urgent care visits. These visits represent a significant burden for patients and the healthcare system, and often signal that toxicities were not detected or managed early enough.
Under-detection is common. Research shows that treatment toxicities experienced by patients are historically under-detected by clinical staff. Physicians and nurses often miss or underestimate the severity of symptoms patients are experiencing, partly because standard clinic visits do not capture how patients feel between appointments. This monitoring gap can allow manageable symptoms to escalate to crisis-level events requiring emergency care.
Patient-reported data shows promise. Patient-reported outcomes (PROs) - standardized questionnaires where patients rate their own symptoms and quality of life - have been shown to improve toxicity detection and clinical outcomes. Studies have found that routine PRO monitoring is associated with better overall survival compared to standard care. The FDA recommends PROs for clinical trials, and clinical guidelines increasingly support their routine use.
Wearables add a new dimension. Wearable sensors like Fitbits capture continuous real-world data on physical activity, sleep, and heart rate - metrics that reflect how patients are actually functioning day to day. A 2023 systematic review found moderate-to-strong evidence linking wearable-derived activity patterns to cancer patient survival, suggesting these devices capture clinically meaningful signals beyond what clinic visits reveal.
Pilot study design. This pilot study enrolled NSCLC patients at Moffitt Cancer Center between March 2021 and March 2022. Eligible patients were starting systemic therapy, had adequate functional status (ECOG performance status 0-2), and were able to provide informed consent. Patients were monitored from before systemic therapy initiation through 60 days of treatment.
Wearable sensor data. Each participant wore a Fitbit device throughout the 60-day study period to continuously capture wearable sensor data (WSD). This included daily physical activity levels, sleep patterns (duration and efficiency), and heart rate. Wearables were chosen because they can capture how patients are functioning between clinic visits, not just at scheduled appointments.
Patient-reported outcome measures. At enrollment and at day 60, participants completed the PROMIS-57 questionnaire - a validated 57-item measure assessing quality of life across seven domains: depression, anxiety, physical function, pain interference, fatigue, sleep disturbance, and social function. This captured subjective wellbeing from the patient's perspective using standardized, validated tools.
Clinical data integration. Medical record data was systematically abstracted, including demographic information and clinical variables such as albumin and calcium levels, treatment types received (chemotherapy, immunotherapy, targeted therapy, radiotherapy), and records of all urgent care visits. Changes in laboratory values from before treatment initiation to the time of urgent care visits were also calculated.
Why Bayesian Networks? Most high-performing machine learning models are "black boxes" - they make predictions but cannot explain why. The researchers deliberately chose Bayesian Networks (BNs) for their explainability. BNs represent relationships among variables as a directed graph where each node is a clinical variable and each edge represents a conditional probability relationship, allowing clinicians to understand the causal pathways driving predictions.
Human-in-the-loop approach. Rather than letting the algorithm build models entirely autonomously, the team used a human-in-the-loop approach that integrates expert clinical knowledge with machine learning. Feature importance rankings from high-performing but opaque methods (random forests, gradient boosting) were used to identify candidates, while interpretable BN models refined the final selection - combining the strengths of both approaches.
Iterative model building. Models were built in an iterative forward stepwise process. The top-ranked features were combined with the BN's Markov blanket (the minimal set of variables needed to predict the outcome) to build an initial model. New features were then added from subsequent importance-ranked intervals, with each new model accepted only if it improved performance. This stops overfitting by adding complexity only when it helps.
Four sequential models. The team built four progressively richer models: a pre-treatment model using only demographic and clinical data; a model adding during-treatment clinical data; a model adding pre-treatment PRO data; and a final model adding wearable sensor data collected during treatment. This sequential design allowed direct quantification of how much each data source improved prediction.
Patient characteristics. The 58 enrolled patients had an average age of 69 years. Most were female (57%), White (88%), and non-Hispanic (95%). Adenocarcinoma was the most common diagnosis (69%). Notably, 69% of patients had at least one urgent care visit during the study period - a high rate that underscores the severity of toxicity burden in this population and the clinical importance of prediction.
Baseline clinical model. The first model using only pre-treatment demographic and clinical data (age, sex, marital status, BMI, cancer stage, adenocarcinoma diagnosis, albumin, calcium) achieved an AUC of 0.72. This is considered moderate predictive accuracy - better than chance but insufficient for reliable clinical decision-making. It establishes the baseline achievable from routinely collected clinical information alone.
Adding treatment data and PROs improves performance. The second model incorporating during-treatment clinical data improved to AUC 0.81. The third model, which added pre-treatment PRO data (global health, anxiety, pain interference scores, age, BMI, functional status), achieved AUC 0.78 on pre-treatment data alone. Both significantly outperformed the first baseline model.
Wearable data achieves best performance. The fourth and final model, incorporating during-treatment clinical data together with wearable sensor data - specifically changes in heart rate and sleep efficiency from before treatment to the time of the urgent care visit - achieved the highest AUC of 0.86 (95% CI: 0.76-0.95). This was significantly better than the corresponding model without wearable data (p less than 0.001), demonstrating the unique informational value of continuous sensor monitoring.
Clinical factors that matter. The Bayesian Network models identified age, albumin levels, calcium levels, and treatment type as important clinical predictors of urgent care visits. Albumin is a blood protein that reflects nutritional status and physiological stress; declining albumin during treatment is a known marker of deteriorating health. Changes in these laboratory values from before treatment to the time of urgent care visits were particularly informative.
PROs that predict visits. Among patient-reported outcomes, global health scores, pain interference, and anxiety from the PROMIS questionnaire emerged as significant predictors. Patients who reported worse overall health status, more pain, and higher anxiety before starting treatment were at greater risk for urgent care visits. These subjective patient assessments captured risk information not reflected in objective clinical metrics.
Wearables reveal hidden deterioration. The wearable sensor data analysis found that changes in heart rate and sleep efficiency from before treatment to the time of urgent care visits were key predictors. Declining sleep efficiency and changes in resting heart rate appear to capture physiological deterioration occurring before patients reach a crisis point requiring urgent care. These signals are invisible to standard clinical monitoring.
Transparent causal pathways. A major advantage of the Bayesian Network approach is that it produces visible relationship maps between variables. The model shows, for example, how treatment type influences laboratory value trajectories, which in turn influence urgent care probability. This transparency allows clinicians to understand the model's reasoning and builds the trust necessary for clinical adoption.
From reactive to proactive care. The primary clinical promise of this approach is shifting cancer care from reactive to proactive. Instead of waiting for patients to become sick enough to seek urgent care, clinicians could use continuous monitoring data to identify patients whose trajectories indicate rising risk, enabling earlier intervention before a crisis develops.
Treatment dosing decisions. The authors suggest that wearable sensor data combined with standard clinical measures could inform treatment dosing adjustments. If a patient's heart rate patterns and sleep quality are worsening, this might signal that the current treatment regimen is creating excessive toxicity burden, prompting a dose modification before the situation escalates to an urgent care visit.
Personalized risk stratification. The Bayesian Network tool could function as a real-time dashboard for oncologists. The study demonstrated how the final model can be used in a clinical interface (using Netica software) to estimate a specific patient's probability of an urgent care visit based on their current demographic, clinical, and wearable data - enabling individualized rather than population-average risk assessment.
Quality and efficiency benefits. Preventing unnecessary urgent care visits benefits patients by sparing them emergency department encounters during an already difficult period, and benefits healthcare systems by reducing costly unplanned care. The ability to identify high-risk patients early could also allow targeted nursing interventions - such as proactive phone calls or earlier clinic appointments - for those most likely to need help.
Proof of concept established. This pilot study successfully demonstrated that combining clinical data with patient-reported outcomes and wearable sensor data in a Bayesian Network framework can predict urgent care visits in NSCLC patients undergoing systemic therapy with meaningful accuracy (AUC 0.86). The study validates the approach and the value of each data source.
Wearable data adds unique value. A key finding is that wearable sensor data provided a significant performance boost that was not achievable with clinical or PRO data alone. This supports expanding wearable monitoring in clinical oncology practice and suggests that the continuous, objective data captured by consumer wearables contains prognostically relevant physiological information.
Important limitations. The study was conducted at a single center with only 58 patients, limiting statistical power and generalizability. The models were not validated in an independent external cohort. The dataset does not include potentially important predictors such as tumor mutation status, vitamin D levels, C-reactive protein, or lactate dehydrogenase. Results should be considered preliminary pending larger validation studies.
Path forward. Future work should extend this approach to larger multicenter datasets, include additional biomarkers and molecular data, develop models for other outcomes (such as treatment delays and overall survival), and extend monitoring beyond the initial 60 days. Validated at scale, tools like this could form the backbone of intelligent, continuous care management systems for cancer patients across the treatment continuum.