Can machine learning predict and prevent mental health problems?

In a current examine printed within the Nature Medicine journal, researchers assessed the efficacy of the machine learning mannequin within the prediction of mental health crises.

The well timed recognition of people prone to mental health deterioration has been reported to enhance illness outcomes. However, manually reviewing affected person information to establish such people is impractical as a result of excessive variety of mental health issues prevalent in society. 

Study: Machine learning mannequin to predict mental health crises from digital health information. ​​​​​​​Image Credit: archy13 / Shutterstock

About the examine

In the current examine, researchers developed a machine learning mannequin that utilized digital health information of people to surveil them for any threat of a mental health disaster for a period of 28 days.

The workforce developed a predictive software to assist healthcare employees streamline caseload priorities whereas enabling skilled intervention that might cut back the danger of a disaster. This was achieved by aligning the prediction goal with the service-oriented strategy with a purpose to outline a disaster as a symptom onset that wanted the intervention of healthcare assets. The workforce generated a dataset that included disaster occasions that had been registered every time a affected person urgently required mental health disaster providers, together with inpatient admission, emergency evaluation of the affected person, dwelling therapy evaluation, or admission to the hospital.

Furthermore, the prediction goal was outlined as the start of a mental health disaster episode, comprising a number of disaster occasions that had been preceded by a minimal of 1 entire steady week with no prevalence of any disaster occasion. Accordingly, the workforce configured the machine learning mannequin to forecast the onset of the primary disaster occasion in an episode throughout the period of the next 28 days.

Time sequence of occasions are represented with the timestamps and occasion traits in numerous SQL tables within the hospital’s database. These tables are processed and transformed into options for the modeling process. Models are skilled, tuned and chosen based mostly on the info for the interval 2012–2019. The system predicts the danger of disaster onset throughout the subsequent 28 days (whereby the algorithm is queried each week for each affected person). The sufferers with the very best predicted threat are displayed on the dashboard delivered to clinicians alongside key indicators, affected person notes and a questionnaire type about every affected person, which the clinician fills out. The icons on this determine had been made by Freepik from IP, inpatient.

The workforce designed the machine learning mannequin such that it allowed queries each week to find out the danger of every affected person experiencing a mental health disaster episode within the subsequent 28 days. The mannequin was constructed by extracting data belonging to 3 classes: (1) static or semi-static affected person knowledge, together with data resembling intercourse, age, and International Classification of Diseases 10 (ICD-10)-coded diagnoses; (2) assessments and interactions obtainable with the hospital (3) variables that represented the time period for the reason that registered occasions. The mannequin generated a predicted threat rating (PRS) between 0 and 1 for every topic.

Machine learning methods resembling choice timber, ensembles, probabilistic, and deep learning-based classifiers had been examined. The workforce subsequently benchmarked the XGBoost (eXtreme gradient boosting) mannequin in opposition to two baseline elements, specifically the clinical-practice- diagnosis-based baseline mannequin. In addition, the prediction mannequin’s efficiency was evaluated amongst sufferers identified with mental health problems categorized in response to the primary degree of the ICD-10.

Furthermore, the workforce carried out a potential examine during which machine learning fashions predicted mental disaster occasions that had been delivered each two weeks to working towards clinicians. This was carried out by querying the mannequin to arrange sufferers in descending order in response to the affected person’s PRS.

Four consultant drive plots, depicting how the options contributed to the prediction for 4 particular knowledge factors. From prime to backside: Patient not going to have a disaster in the course of the subsequent 4 weeks (goal=0), the mannequin assigned a prediction worth of 0.178. Patient not going to have a disaster in the course of the subsequent 4 weeks (goal=0), the mannequin assigned a prediction worth of 0.129. Patient going to have a disaster in the course of the subsequent 4 weeks (goal=1) the mannequin assigned a prediction worth of 0.792. Patient going to have a disaster in the course of the subsequent 4 weeks (goal=1) the mannequin assigned a prediction worth of 0.725.


The examine cohort included 5,816,586 information that had been obtained from 17,122 particular person sufferers aged between 16 and 102 years within the interval between September 2012 and November 2018. Patients within the group had been identified with quite a few problems, together with temper, natural, neurotic, psychotic, and character problems. The knowledge set additionally included 60,388 disaster episodes, with a median of 24 disaster occasions per episode.

The examine outcomes confirmed that the overall mannequin carried out considerably higher within the case of natural problems with an space below the receiver working attribute (AUROC) of 0.89 compared to the general efficiency of 0.797. When different diagnostic teams had been taken into consideration, the overall and total efficiency was between 0.770 and 0.814, respectively. The workforce noticed that the bottom efficiency was for temper problems, schizophrenia, and schizotypal and delusional problems. In the comparability of separate fashions with the overall one, the overall mannequin carried out higher than the baseline fashions whereas not one of the disorder-specific fashions displayed higher efficiency than the overall mannequin.

The workforce additionally assessed the efficiency of the overall mannequin by way of affected person age teams. For sufferers aged lower than 18 years, the overall mannequin efficiency decreased to 0.743, whereas for these aged between 65 and 74 years, the efficiency rose to 0.840. Notably, comparable efficiency was noticed for the sufferers belonging to different age teams with the AUROC ranging between 0.782 and 0.796.   

The examine exhibits that machine learning can present a basis for facilitating higher useful resource allocation in mental healthcare. Researchers imagine the current examine can encourage mental health suppliers to shift from reactive care to preventative care.

Recommended For You