Application of Multi-Layer Perceptron (MLP) Machine Learning Model in Determining Mental Health Illness Severity

Spikes in mental health (MH) awareness have led to the development of the 988 hotline, which was created to personally assist with MH issues. Hotline operators analyze a patient's specific MH crisis, illnesses, and other provided information, all to make a fast triage judgment on the severity of the patient's condition. Determining condition severity requires holistic consideration of factors (socioeconomic, specific illnesses, etc.) that one cannot instantaneously process; however, machine learning (ML) models can expedite the analysis. This research explores Keras’s Sequential multi-layer perceptron neural network model’s use in determining severe mental illness (SMI) in patients (i.e., illness that substantially interferes with a person's life and function). Using client-level data from SAMHSA's (Substance Abuse and Mental Health Services Administration) 2023 dataset, the model was trained on over 50,000 adult patients to predict the SMI target field (serious or not serious). Using socio-economic features (age, race, and gender) and known MH conditions of patients, the model accuracy is 85.44% in concluding the presence of SMI in a patient, with 92.53% precision and 85.50% recall in determining SMI patients and 73.47% precision and 85.32% recall in determining non-SMI patients. This study also considers factor importance when determining a patient's SMI, removing certain indicators to determine the impact on model performance. For instance, the removal of the patient's MH1 (first diagnosed MH condition) in consideration tanks model accuracy down to 78% with 89.01% precision and 76.42% recall in determining SMI patients and 61.47% precision and 79.94% recall in determining non-SMI patients. While findings don’t support artificial intelligence (AI) full replacement of 988 operators, they show that AI models can have value as tools for operators in assisting an accurate assessment of a patient’s mental state efficiently. For future improvement of the ML model, datasets with anecdotal behavior descriptions will be used in training for increased model accuracy.