- Title: “Objective Assessment of Psychiatric Patients via Machine Learning”
- Advisor: Antonio Artés Rodríguez.
Mental disorders are still a source of not-well-understood human suffering. They affect one out of four people in the world, and they are more costly to treat than cancer and diabetes together. One of the reasons of the high economic burden is the lack of an objective tool to continuously assess the health condition of the patients. Hitherto, psychiatrists mainly rely on interviews at clinical sites where the patients usually talk about their recent, yet past events. Those evaluations are therefore biased by both the memory of the patients and the fact that they must go outside their usual environment. This method misses opportunities to prevent relapses, particularly between consultations, when the psychiatrists are blind to the current situation of the patients. Hence, if the patients’ health condition worsens, yet, they do not try to look for help, they may eventually suffer a relapse. Then, they may need to be hospitalized, what involves high economical and human burdens.
This thesis represents a step towards the overarching aim of the objective evaluation of psychiatric patients. In order to do so, two approaches are explored. On the one hand, new algorithms are developed to automatically extract meaningful and interpretable patterns from the social interactions of the patients via their electronic devices. This is done without requesting any particular intervention from the patients. The features that are extracted from those interactions (known as digital phenotype) include: the times at which the patients make phone calls, their durations, anonymous identifiers of the callers, the call types, and the times at which the patients use social and communication apps.
In particular, the event times are modeled by a family of Poisson point processes that uses non-negative Fourier series to capture the circadian (i.e., daily) rhythm present in human behavior. The modeling capabilities are extended to account for different daily profiles and for simultaneous, heterogeneous sources of information. In addition, the outgoing probability profile, which can be used to see if a patient takes the initiative to start a conversation, is obtained by a marked Poisson process. The switching Poisson process is used to model phone call durations. This process makes use of two intertwined, yet independent intensities: one that models the beginning, and other one that models the end of calls. Finally, a mixture of exponential distributions is used to cluster callers based on their durations.
On the other hand, this thesis explores a different approach in which the patients are asked to answer, with the help of their electronic devices, a set of health-related questions while they perform their usual activities. Since the variability of suicidal thoughts and other clinical factors during follow-up has emerged as a promising phenotype to identify vulnerable patients, the analysis is focused on the variability of suicidal patients. To that end, several variability metrics are compared and proposed. Then, the most reliable and interpretable metric is used to analyze a cohort of suicidal patients. It is found that suicidal patients are best clustered in two groups depending on their variability: the high- and low-variability groups. Both groups are separated by ten clinical features, including depressive symptoms, cognitive instability, the intensity and frequency of suicidal ideation, and the occurrence of clinical events, such as suicide attempts or emergency visits during follow-up.