A Feasibility Study of an AI-Powered Clinical Decision Aid for Personalized Depression Treatment Selection

A Feasibility Study of an AI-Powered Clinical Decision Aid for Personalized Depression Treatment Selection

A Feasibility Study of a Hybrid-Classic/Deep-Learning Enabled Clinical Decision Aid for Personalized and Individualized Pharmacological Depression Treatment Selection

The Clinical Decision Aid (CDA) is a predictive model that takes as input individual patient characteristics, called 'features', which are inputted by the physician or by patient self-report, and outputs a list of possible treatments, with each treatment associated with a predicted efficacy (likelihood to achieve response and likelihood to achieve remission, each expressed as a percentage). The treatments, which may include any approved treatment for depression, will be presented to the physician who will then make a treatment choice.

Hypothesis 1. There will not be a significant difference in measured non-initial intake appointment lengths between the baseline period and the appointment length measured at two and four months after introduction of the study software and CDA. Hypothesis 2. Physicians will not subjectively report that using the CDA and study software increased the length of their appointments. Hypothesis 3. At least 66% of patients and 66% of physicians will rate the trustworthiness of the CDA as a 4 or 5 on a 5 point Likert scale (with higher ratings indicating greater trust). Hypothesis 4. At least 66% of patients and 66% of physicians will rate the overall usability of the CDA as a 4 or 5 on a 5 point Likert scale (with higher ratings indicating greater usability). Hypothesis 5. At least 70% of physicians and 65% of patients will still be using the application regularly by the end of the study. For physicians, regularly will be defined as the application being used in every study-related visit. For patients regularly will be defined as completing at least one PHQ-9 and GAD-7 questionnaire on the application per week. Exploratory hypothesis: Based on our machine learning results to date, we expect between 40-50% of patients starting a new treatment for depression and whose treatment follows the highest probability treatment output by the CDA to remit within 14 weeks. This is exploratory, and the study is not necessarily powered to demonstrate this.

The Clinical Decision Aid is a predictive model that takes as input individual patient characteristics, called 'features', which are inputted by the physician or by patient self-report, and outputs a list of all possible treatments, with each treatment associated with a predicted efficacy (likelihood to achieve response and likelihood to achieve remission, each expressed as a percentage). The treatments, which may include any approved treatment for depression, will be ordered by efficacy and presented to the physician. Lifestyle interventions, such as exercise or mindfulness, which have an evidence base, but do not require formal regulatory approval, will also be outputted. The system will additionally produce a side effect profile for each pharmacological treatment recommended, including known side effects, modified by a prediction about which side effects may be more likely for a given individual based on their individual characteristics.

We will be using our Clinical Decision Aid Feasibility Questionnaire (Version 1), a descriptive questionnaire with 5-point Likert scales (with higher values representing better outcomes) and narrative questions about experience using the tool.

Inclusion Criteria: All patients of the physicians in the study are diagnosed with major depressive disorder by a physician using DSM-V criteria. All participants must be able to provide informed consent. Contraception will be used as per established clinical guidelines and usual clinical practice for medications known to cause birth defects. The medications prescribed and the use of and type of contraception will be determined by the physicians in the study in consultation with their patients as would usually occur in clinical practice. Exclusion Criteria: Bipolar disorder type 1 or 2, as the data we have used to train the model does not allow for generalization to bipolar disorder (either pre-existing or as diagnosed according to DSM-5 criteria). Inability or unwillingness of individual to give informed consent.

Popescu C, Golden G, Benrimoh D, Tanguay-Sela M, Slowey D, Lundrigan E, Williams J, Desormeau B, Kardani D, Perez T, Rollins C, Israel S, Perlman K, Armstrong C, Baxter J, Whitmore K, Fradette MJ, Felcarek-Hope K, Soufi G, Fratila R, Mehltretter J, Looper K, Steiner W, Rej S, Karp JF, Heller K, Parikh SV, McGuire-Snieckus R, Ferrari M, Margolese H, Turecki G. Evaluating the Clinical Feasibility of an Artificial Intelligence-Powered, Web-Based Clinical Decision Support System for the Treatment of Depression in Adults: Longitudinal Feasibility Study. JMIR Form Res. 2021 Oct 25;5(10):e31862. doi: 10.2196/31862.