JMIR Publications

ABSTRACT

Background:

Machine learning (ML) systems are part of Artificial Intelligence (AI) that automatically learn models from data in order to make better decisions. Natural Language Processing (NLP), by using corpora and learning approaches, provides good performances in tasks of statistical nature, like text classification or sentiment mining.

Objective:

The first aim of this systematic review is to summarize and characterize studies that used ML and NLP techniques for mental health, in methodological and technical terms. The second is to consider the interest of these methods in the mental health clinical practice.

Methods:

This systematic review follows the PRISMA guidelines and is registered on PROSPERO (number CRD42019107376). The research was conducted on 4 medical databases (Pubmed, Scopus, ScienceDirect and PsycINFO) with the following keywords: machine learning, data mining, psychiatry, mental health, mental disorder. The exclusion criteria are: other languages than English, anonymization process, case studies, conference papers and reviews. No limitations on publication dates were imposed.

Results:

327 articles were identified, 269 were excluded and 58 included in the review. Results are organized through a qualitative perspective; even though studies have heterogeneous topics and methods, some features emerged. Population studies can be grouped in three categories: patients included in medical databases, patients who came to the emergency room, social-media users. Main objectives are symptom extraction, severity of illness classification, comparison of therapy effectiveness, psychopathological clues, and nosography challenging. Data from electronic medical records and that from social media are the two major data sources. Concerning the methods: preprocessing used the standard methods of NLP and unique identifier extraction dedicated to medical texts; efficient classifiers are preferred rather than "transparent" functioning ones; Python is the most frequently used platform.

Conclusions:

ML and NLP models have been a highly topical issue in medicine for recent years and may be considered to be a new paradigm in medical research. However, these processes tend to confirm clinical hypotheses rather than developing entire new knowledge, and one major category of population is social-media users, which is obviously an imprecise cohort. In addition, some language-specific features can improve the performance of NLP methods, and their extension to other languages should be more investigated. However, ML and NLP techniques provide information from unexplored data, i.e. patient's daily habits that are usually inaccessible to care providers. It may be considered to be an additional tool at every step of mental health care: diagnosis, prognosis, treatment efficacy and monitoring. Therefore, ethical issues – like predicting psychiatric troubles or involvement in the physician-patient relationship – remain and should be discussed in a timely manner. ML and NLP methods may offer multiple perspectives in mental health research but should be considered to be a tool to support clinical practice. Clinical Trial: Registered on PROSPERO (number CRD42019107376), updated on 17th of March 2020