Table 5.

Comparison of this and prior studies on coercive measures employing machine learning.

	(30)	(15)	Current study
Topic of study	Predictors for direct coercive measures in patients with all diagnoses in general psychiatry	Predictors for mechanical restraint in patients with all diagnoses in general psychiatry	Predictors for direct coercive measures in patients with schizophrenia in forensic psychiatry
Sample studied	Patients with coercion: 170	Patients with mechanical restraint: 5050	Patients with coercion: 131
Data collection	Retrospective file content analysis	Retrospective health record and registry content analysis	Retrospective file content analysis
Number of potential predictors explored	Not specified	86	569
Similar predictor variables at statistical significance	Threat of violence as reason for involuntary admission1, prior involuntary admission to treatment, antipsychotic medication	Threat of violence measured with the Broset violence checklist, involuntary admission to treatment, threatening/abnormal behavior, sparse/non-coherent/non-informative verbal response	Threat of violence, coercive measures in prior treatment(s), haloperidol prescribed, daily olanzapine equivalent prescribed upon discharge, poor impulse control, hostility and uncooperativeness at admission, total PANSS-score at admission
Model accuracy (balanced)	66.5–78.5%	Not specified	73.3%
ROC AUC	0.73–0.75	0.87	0.8468
Sensitivity	60–69%	56%	72.87%
Specificity	78–83%	94%	73.68%

ROC AUC, receiver operating characteristic curve area under the curve method, a measure for the goodness of fit of a model (67); PANSS, Positive and Negative Symptom Scale.

¹Authors see a limitation in their measuring threat of violence only in terms of reason for involuntary admission.