When it comes to understanding bipolar and mental disorders, a wealth of relevant data can be obtained from a smartphone, as it’s what most people interact with constantly on a daily basis.
A new study from the Center for Research and Telecommunication Experimentation for Networked Communities (CREATE-NET) in Trento, Italy, explores how analysing data on a bipolar patient’s daily activity collected from smartphone sensors such as GPS and voice calls could help predict the onset of an episode.
Venet Osmani, from CREATE-NET, who conducted the study, monitored 12 patients’ daily activities, each for 12 weeks on average. He collected more than 1000 days of smartphone sensor data. The study was approved by the ethics board of the Innsbruck University Hospital in Austria.
“Behavioural data will have a significant impact on our understanding of mental disorders. Because symptoms of most mental disorders are manifested as changes in an individual’s behaviour, analysing such changes could lead to a better understanding of these types of diseases and possible treatments,” he wrote in his research paper.
“Detecting patients’ change of state (which can indicate onset of an episode) can lead to a visit to the clinic and allow early intervention.”
Osmani said current clinical rating scales for diagnosing bipolar disorders - such as Hamilton Depression Rating Scale (HAMD) and Bipolar Spectrum Diagnostic Scale (BSDS) - have their drawbacks as they tend to be subjective. The study aims to address this issue through more objective data analysis to better understand bipolar disorders and the behaviour changes that lead to episodes.
Patients in the study underwent a psychiatric mental state examination every three weeks, including one at the beginning and one at the end. This was used for ‘ground truth’ data to compare the predicted result against the actual situation. The predictive models were assessed against whether the examination gave a patient a score that states either an episode of severe depression, an episode of severe mania, or moderate and mild conditions.
“We chose a period of seven days before and two days after the mental state examination for the sensor data. This was based on the assumptions elicited from discussions with the psychiatrists that state changes are gradual and the probability of a major change within a few days is low.”
Osmani first looked at the correlation between a patient’s activity and their mental state using data collected from a smartphone accelerometer. Dividing the day into morning, afternoon, evening and night, he calculated an activity score for each part of the day. A strong correlation was established between a patient’s activity and their mental state at particular parts of the day.
GPS data were used to train a Naïve Bayes classification model to predict a patient’s mental state, which achieved 81 per cent mean accuracy when tested against the ground truth. Osmani also tried k-Nearest Neighbour, J48 search tree, and a conjunctive rule learner yielded - all giving similar accuracy results.
Osmani then included voice phone call patterns and sound analysis, as bipolar patients experience voice changes during an episode. By fusing all the sensor data together and considering all disease-relevant aspects of behaviour, he was able to achieve a prediction accuracy of 97.19 (percentage of correct positive predictions) and of 97.36 per cent for recall (how often model found positives).
“We plan to investigate whether similar results can be obtained with a higher number of patients, monitored over a longer period of time," said Osmani.
The research team has begun the initial steps in this direction through the Nympha-MD (www.nympha-md-project.eu), a follow-up European project.