Predicting Resource Use of Community Mental Health Services at the Transition from Inpatient Psychiatry

Mental health is a major health problem for many Canadians. Methods to predict expected mental health care resource use are an essential component in balancing the needs of the population and equitable allocation of limited health care resources. This research examined the relationship between the resource use of community mental health services and the characteristics of their clients using a case-mix classification approach. \n \nA scoping review showed that most of the research on this topic focused on inpatient psychiatry settings. The number of identified studies (n=17) and case-mix systems (n=32) reflected the modest level of research activity in this area. \n \nSecondary analyses were done with a sample of adults discharged from a local psychiatric hospital unit in Ontario (n=4,688 discharges) that was tracked to examine the use of community mental health services after discharge. Only about half of the discharges subsequently used publicly funded community mental health services. Further, only n=1,207 discharges had services initiated within 30 days and were not censored by readmission. Clinical characteristics measured at discharge from inpatient psychiatry were associated with observed use and high use (as binary variables) of community mental health services post-discharge. Usage of services specially designed for persons at risk of self-harm and harm to others (as binary variables) were also associated with higher risk of self-harm and harm to others measured at discharge. \n \nA community episode of 90 days from first contact with the community mental health agency post-discharge appeared to be the most practical for implementation. Two high performing case-mix classification systems were examined for their possible predictive utility for post-discharge community mental health service use. The System for Classification of In-Patient Psychiatry (SCIPP) achieved 6% explained variance of community resource use for an episode. When prior contact with the community mental health agency within 30 days prior to the inpatient episode was included, the model with SCIPP explained up to 14.1% of variance in resource use. The Australian Mental Health Classification (AMHCC) was found to be not immediately applicable outside of the Australian context, and most of its explained variance was likely attributed to the “phases of care” that are subjectively determined by clinicians at the beginning of an episode. The remaining components of the AMHCC explained only 1.2% of variance in resource use. \n \nUsing machine learning, new classification models using discharge clinical characteristics achieved up to 12% of explained variance in cross-validation. The two simplest decision tree models showed similar performance in cross-validation as more complex models. Although machine learning identified relevant relationships between clinical characteristics and observed resource use, some relationships required human expertise to adjust to align with the goals of the health care system. This was exemplified by a manual decision tree model that achieved 11.1% explained variance on the development data set. These results pointed to the need for additional research to: expand the sample size; include a broader range of community mental health service users; use more contemporaneous clinical assessment data measured at community service initiation; and broaden the participation of community mental health agencies. Although clinical characteristics measured at discharge yielded only modest predictive utility, designing a system that could leverage both inpatient information and community agency assessment information could improve both predictive utility and care integration across the care continuum. Further development of case-mix classification for community mental health will require a broad collaboration across the health care system.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.