Investigating the use of Artificial/Intelligence/Machine Learning for early screening of mental health disorders using primary care data

Lead Research Organisation: University of Warwick
Department Name: WMG

Abstract

Using data available within primary care/public area such as Electronic Health Records/Social Media establish if it is possible using a combination of statistical methods and machine learning to provide reliable prediction of disorder development. This could then be used to "red flag" to primary care practitioners that further diagnosis/screening is required to determine if early intervention is beneficial. This builds on earlier work at WMS by Nichols et al. (2016) on depression in children/young adults. The first step would be to replicate this and see if AI/ML could deliver improved specificity/reliability before considering other disorders, age groups and methods. A sub project would be establishing what such an application would need to achieve to be acceptable to primary care staff, possibly via focus groups.
REFERENCE: Nichols, L., Ryan, R., Connor, C., Birchwood, M. and Marshall, T. (2018) 'Derivation of a prediction model for a diagnosis of depression in young adults: a matched case-control study using electronic primary care records', Early Intervention in Psychiatry, vol. 12, no. 3, pp. 444-455 [Online]. DOI: 10.1111/eip.12332.
Alligns with the EPSRC research area in Clinical Technologies (excluding imaging)

Studentship Projects

Project Reference Relationship Related To Start End Student Name
EP/R513374/1 01/10/2018 30/09/2023
2300953 Studentship EP/R513374/1 30/09/2019 30/03/2023 David Nickson
 
Description Demonstrated the potential to develop a viable machine learning implementation to support primary care practitioners in diagnosing depression. This should offer a good level of performance in terms of both sensitivity and specificity, offering a reasonable balance between false negatives with their potential for poorer outcomes and false positive results with their potential for wasting resources. Identified key issue of acceptability for practitioners and how information might be presented as possible future work.
I have replicated Nichols et al. (2018) study using a new non overlapping data set and this has shown an improved performance that is possibly due to increased presentation of mental health conditions, many of which are strong predictors of depression, in the time that has passed since the original data set was collected. In addition to the original logistic regression based models alternatives models have been developed using, Prediction Rules Ensembles, XGBoost, Gradient Boosting, LASSO (Least Absolute Shrinkage Selection Operator), Rpart and Random Forest. These models delivered comparable performance, sometimes marginally better/worse than for the original Logistic Regression (LR) models. The available presentation of data from these models is being assessed in terms of acceptability for practitioners to deliver a depression diagnosis service to their patients. The requirement to anonymize the data restricts the data available for predictions. For example, social deprivation data was only available at practice level, a granularity placing thousands of people in the same quintile. It is known that this is a good predictor at an individual level, so this restriction is likely to reduce the overall accuracy of the models.

A systematic review has also been completed and was prospectively registered with Prospero, the international database of systematic reviews (# CRD42021269270) as "Predicting depression using Electronic Health Records data: a systematic review." Both the main project and the systematic review are in preparation as journal papers for submission during 2022.
Exploitation Route Two areas for further work have been identified. The first is to conduct a trial of such an application in a real primary care environment, one in which the data did not have to be fully anonymized to establish performance in a real world environment. The second is a qualitative study to explore acceptability of such a system in primary care for health care practitioners and patients. These could be explored by the current team or others subject to funding.
Sectors Digital/Communication/Information Technologies (including Software),Healthcare