Part 3 – AI use in relation to specific concerns

Article in Review: D’Hotman, D. & Loh, E. (2020). AI-enabled suicide prediction tools: a qualitative narrative review. BMJ Health Care Inform, doi:10.1136/bmjhci-2020-100175

Summary: This article is the third of a four-part series examining the article by D’Hotman & Loh (2020) and their narrative review of artificial intelligence (AI) examining the capability of using AI for suicide prediction via data gathered through medical records and social media. The first two articles examined some of the challenges in the area, the use of data from medical sources (such as medical records), and the social media content. In the current article, we review D’Hotman’s & Loh’s (2020) discussion of mental and physical health conditions, in the context of potential suicide prediction via AI. To be honest, although their observations are startling, they are also kind of cool. Don’t miss the fourth and final article when we review their observations of Facebook being used specifically as a ‘suicide prevention tool’.

___________________________________________________________________

There are multiple AI data tools and techniques that can be used to assess large data sets. D’Hotman and Loh (2020) have reviewed a number of studies involving both large and small populations and data sets to demonstrate the broad and beneficial range of applications that AI in data might have for clinicians assessing patients at risk of, or who are suicidal. Some of these relevant studies are summarised in this third blog related to the article AI-enabled suicide prediction tools: a qualitative narrative review.

While only a small fraction of people with mental illness will die by suicide, more than 80% of those who do are thought to have a mental illness. We know that risk increased in soft rations with multiple mental illnesses and often mental illnesses overlap. At least 50% of patients receive more than one psychiatric diagnosis. There is interest in understanding the risk of mental illness in the instances of people who may also be at risk of suicide.

Depression, anxiety, and mood disturbances

• An AI model has shown the capacity to identify depressed individuals based on speaking patterns, with range and pitch and the number of stops, pauses and starts between words.
• An AI model could identify patients with anxiety and depression and even the severity of the illness based on their walking style.
• Machine assessment of 44,000 Instagram photos identified subjects who had markers of depression with 70% accuracy.

Schizophrenia

• The use of alterations in brain activity captured by MRI imaged could correctly diagnose schizophrenia with 87% accuracy when the chance rate is 53%.

Post-traumatic stress disorders

• Machine learning was used to analyse risk indictors and predict long-term post-traumatic stress responses in a group of Danish soldiers who were tracked for six years. the results indicated potentially significant benefits in identifying high-risk soldiers and improving treatments

Suicide amongst adolescences

• In a sample of nearly 6,0000 Korean adolescents, five AI techniques were used to determine the risk of suicide through analysis of data related to the history of suicide attempts and ideation. The study achieved close to an 80% accuracy rate
• Another study looked used AI to look at 50,0000 medical records of young Californians to predict suicide attempts. The strongest performing model in the experiment achieved a sensitivity of 70% and specificity of 98%.

Suicide ideation

• Suicide ideation can be difficult to recognise, because many people may not be willing to disclose it to their doctors. Tools that can predict suicide ideation based on psychological factors might be valuable, especially for high-risk groups like veterans.
• A study from Taiwan that used machine learning to direct suicidal ideation in men and women from the military. It was determined to achieve an accuracy of over 98% in predicting suicide ideation. It was shown to improve the sensitivity of critical criterion by 35% and precision by 65%
• Suicide ideation amongst Reddit users was assessed using long short-term memory and convolutional neural networks with an accuracy rate of 93%. Statistical, linguistic, topical, and word embedding techniques were used in another study which had a 90% accuracy in identifying suicide ideation among Reddit and Twitter users.
• Tools to assess suicidal ideation might be even more effective when combined with data and knowledge related to suicide attempts and completed suicides, enabling us to provide targeted care.

Self-injury and self-harm

• One study here used data to determine that suicide plans and depression, both suicide risk factors were also indicators of lifetime self-harm risk.
• Another used machine learning technique to determine motivation, methods lethality and scarring, and the most important factors for ascertaining suicide risk.

Physical illness

• The presence of a physical illness has previously been found to contribute to scud risk. Machine learning techniques were used to assess data related to 7,399 mental health patients with a risk of physical illness. The data significantly outperformed clinical baseline risk assessment in predicting the risk of suicide.

When supporting patients with complex situations and multiple medical diagnoses, AI tools might provide better diagnostic clarity for clinicians. Better diagnostics will only improve treatment options and subsequently enable improved access to mental health services for those who are the most vulnerable and at risk.

Part 4 – When Facebook finds out first