Maximizing the Value of Mobile Health Monitoring by Avoiding Redundant Patient Reports: Prediction of Depression-Related Symptoms and Adherence Problems in Automated Health Assessment Services

BACKGROUND: Interactive voice response (IVR) calls enhance health systems’ ability to identify health risk factors, thereby enabling targeted clinical follow-up. However, redundant assessments may increase patient dropout and represent a lost opportunity to collect more clinically useful data. OBJECTIVE: We determined the extent to which previous IVR assessments predicted subsequent responses among patients with depression diagnoses, potentially obviating the need to repeatedly collect the same information. We also evaluated whether frequent (ie, weekly) IVR assessment attempts were significantly more predictive of patients’ subsequent reports than information collected biweekly or monthly. METHODS: Using data from 1050 IVR assessments for 208 patients with depression diagnoses, we examined the predictability of four IVR-reported outcomes: moderate/severe depressive symptoms (score ≥10 on the PHQ-9), fair/poor general health, poor antidepressant adherence, and days in bed due to poor mental health. We used logistic models with training and test samples to predict patients’ IVR responses based on their five most recent weekly, biweekly, and monthly assessment attempts. The marginal benefit of more frequent assessments was evaluated based on Receiver Operator Characteristic (ROC) curves and statistical comparisons of the area under the curves (AUC). RESULTS: Patients’ reports about their depressive symptoms and perceived health status were highly predictable based on prior assessment responses. For models predicting moderate/severe depression, the AUC was 0.91 (95% CI 0.89-0.93) when assuming weekly assessment attempts and only slightly less when assuming biweekly assessments (AUC: 0.89; CI 0.87-0.91) or monthly attempts (AUC: 0.89; CI 0.86-0.91). The AUC for models predicting reports of fair/poor health status was similar when weekly assessments were compared with those occurring biweekly (P value for the difference=.11) or monthly (P=.81). Reports of medication adherence problems and days in bed were somewhat less predictable but also showed small differences between assessments attempted weekly, biweekly, and monthly. CONCLUSIONS: The technical feasibility of gathering high frequency health data via IVR may in some instances exceed the clinical benefit of doing so. Predictive analytics could make data gathering more efficient with negligible loss in effectiveness. In particular, weekly or biweekly depressive symptom reports may provide little marginal information regarding how the person is doing relative to collecting that information monthly. The next generation of automated health assessment services should use data mining techniques to avoid redundant assessments and should gather data at the frequency that maximizes the value of the information collected.