Predicting Adolescent Suicide Attempts with Neural Networks

Suicide is a major public health issue in the United States and machine learning methods have not been widely used to predict an individual's risk of attempting or committing suicide. To address this problem, Harish S. Bhat and Sidra J. Goldman-Mellor developed neural network models to predict suicide attempts among California-resident adolescents using anonymized electronic health records for 522,056 unique patients. The data from 2006-2009 was used to frame the problem as a binary classification task where the presence (1) or absence (0) of a suicide attempt in 2010 was predicted. To tackle issues such as imbalanced classes and variable patient history lengths, neural networks with depths ranging from two to eight hidden layers were built. The depth-4 model achieved a sensitivity of 0.703, specificity of 0.980, and AUC of 0.958 on test set observations where there were at least five ED/hospital visits' worth of data on a patient. This research shows the potential for machine learning methods to be applied in predicting suicide attempts and emphasizes the importance of addressing challenges such as imbalanced classes and variable patient history lengths when developing predictive models for mental health outcomes.

Using Machine Learning to Predict Suicide Attempts Among Adolescents

Suicide is a major public health issue in the United States, yet machine learning methods have not been widely used to predict an individual's risk of attempting or committing suicide. Harish S. Bhat and Sidra J. Goldman-Mellor recently developed neural network models that can accurately predict suicide attempts among California-resident adolescents using anonymized electronic health records (EHR). In this article, we will discuss their research paper and explore how machine learning can be applied to mental health outcomes such as predicting suicide attempts.

Background

The research paper by Bhat and Goldman-Mellor focused on developing predictive models for suicide attempts among California adolescents aged 12–17 years old using EHR data from 2006–2009. The data was framed as a binary classification task where the presence (1) or absence (0) of a suicide attempt in 2010 was predicted. To tackle issues such as imbalanced classes and variable patient history lengths, neural networks with depths ranging from two to eight hidden layers were built.

Methods

The authors used 522,056 unique patients’ EHR data from 2006–2009 to build their predictive models for adolescent suicides in 2010. To address the challenge of imbalanced classes, they employed stratified sampling techniques during model training so that each class had equal representation in the training set. Additionally, they addressed the challenge of variable patient history lengths by creating separate models for different numbers of ED/hospital visits: 0–4 visits; 5+ visits; 10+ visits; 15+ visits; 20+ visits; 25+ visits; 30+ visits; 35+ visits; 40+visits+.

The authors then compared several different neural network architectures including shallow networks with 2 hidden layers and deep networks with 8 hidden layers. They found that depth-4 model achieved the best performance on test set observations where there were at least five ED/hospital visit’s worth of data on a patient: sensitivity = 0.703, specificity = 0.980, AUC = 0.958 .

Discussion

This research shows promising results for applying machine learning methods to predict suicide attempts among adolescents using EHR data from 2006–2009 . It also emphasizes the importance of addressing challenges such as imbalanced classes and variable patient history lengths when developing predictive models for mental health outcomes like predicting suicidal behavior . Further studies should focus on expanding these findings beyond California residents by collecting more diverse datasets across multiple states or countries , as well as exploring other types of machine learning algorithms that could potentially improve prediction accuracy even further .