Background: There is a lack of knowledge regarding the actionable key predictive factors of homelessness in psychiatric populations. Therefore, we used a machine learning model to explore the REHABase database (for rehabilitation database-n = 3416), which is a cohort of users referred to French psychosocial rehabilitation centers in France.
Methods: First, we analyzed whether the different risk factors previously associated with homelessness in mental health were also significant risk factors in the REHABase. In the second step, we used unbiased classification and regression trees to determine the key predictors of homelessness. Post hoc analyses were performed to examine the importance of the predictors and to explore the impact of cognitive factors among the participants.
Results: First, risk factors that were previously found to be associated with homelessness were also significant risk factors in the REHABase. Among all the variables studied with a machine learning approach, the most robust variable in terms of predictive value was the nature of the psychotropic medication (sex/sex relative mean predictor importance: 22.8, σ = 3.4). Post hoc analyses revealed that first-generation antipsychotics (15.61%; p < 0.05 FDR corrected), loxapine (16.57%; p < 0.05 FWER corrected) and hypnotics (17.56%; p < 0.05 FWER corrected) were significantly associated with homelessness. Antidepressant medication was associated with a protective effect against housing deprivation (9.21%; p < 0.05 FWER corrected).
Conclusions: Psychotropic medication was found to be an important predictor of homelessness in our REHABase cohort, particularly loxapine and hypnotics. On the other hand, the putative protective effect of antidepressants confirms the need for systematic screening of depression and anxiety in the homeless population.
Keywords: REHABase; antipsychotics; classification and regression tree model (CART); depression; homelessness; machine learning; psychotropic medication.