Impact factor analysis: combining prediction with parameter ranking to reveal the impact of behavior on health outcome

An increasing number of healthcare systems allow people to monitor behavior and provide feedback on health and wellness. Most applications, however, only offer feedback on behavior in form of visualization and data summaries. This paper presents a different approach--called impact factor analysis--in which machine learning techniques are used to infer the progression of a primary health parameter and then apply parameter ranking to investigate which behavioral data have the highest `impact' on health. We have applied this approach to improve the MONARCA personal health application for patients suffering from bipolar disorder. In the MONARCA system, patients report their daily mood score and by analyzing self-reported and automatically sensed behavioral data with this mood score, the system is able to identify the impact of different behavior on the patient's mood. We report from a study involving ten bipolar patients, in which we were able to estimate mood values with an average mean absolute error of 0.5. This was used to rank the behavior parameters whose variations indicate changes in the mental state. The rankings acquired from our algorithms correspond to the patients' rankings, identifying physical activity and sleep as the highest impact parameters. These results revealed the feasibility of identifying behavioral impact factors. This data analysis motivated us to design an impact factor inference engine as part of the MONARCA system. To our knowledge, this is a novel approach in monitoring and control of mental illness, and we argue that the impact factor analysis can be useful in the design of other health and wellness systems.