With help of Artificial Intelligence, it is now possible to prevent suicide

A quarter of suicide victims in the UK had spoken to a healthcare provider in the previous week, and the majority had done so during the previous month. Yet it is still very challenging to determine the risk of suicide in patients. In England in 2021, 5,219 suicide fatalities were officially documented. The majority of this decrease occurred before 2000, despite the fact that the suicide rate in England and Wales has decreased by about 31% since 1981. Men commit suicide at a rate that is three times higher than that of women, and this difference has grown over time. Numerous studies on artificial intelligence are being conducted in different medical specialties, like cancer. However, despite their promise, healthcare settings have not yet made extensive use of artificial intelligence models for mental health.

The Black Dog Institute at the University of New South Wales led a study in October 2022 that discovered artificial intelligence models beat clinical risk evaluations. It examined 56 studies between 2002 and 2021 and discovered that artificial intelligence accurately predicted 87% of persons who would not commit suicide and 66% of people who would. Traditional scoring techniques used by health experts, in contrast, perform only marginally better than random. According to a 2019 study from Sweden's Karolinska Institute, four conventional scales that were used to gauge the likelihood of suicide following recent self-harm events performed poorly. The difficulty in predicting suicide arises from a patient's aim changing suddenly.

Health practitioners in England are explicitly advised not to rely on suicide risk assessment tools and scales in their guidelines on self-harm. Professionals should instead conduct clinical interviews. While systematic risk assessments are performed by clinicians, they are more often used to enhance interviews than to assign patients to one of several treatment options.

The Black Dog Institute study generated encouraging findings, but if 50 years of research into conventional (non-artificial intelligence) prediction led to techniques that were only marginally better than chance, we need to question whether we can put our faith in AI. It can be tempting to cease asking questions when a new invention delivers us what we desire (in this case, better estimates of the danger of suicide). However, it would be unwise to rush this technology. If you get it wrong, you could literally lose your life. Every AI model has constraints, including the manner in that performance is measured. For instance, if the dataset is uneven, utilizing accuracy as a statistic may be misleading. A model that consistently predicts there won't be a suicide risk and can attain 99% accuracy if only 1% of the patients in the dataset are considered to be at high risk, it is always predicted that there would be no suicide risk.

It's also crucial to evaluate AI models using data that is unrelated to that on which they were trained. In order to prevent overfitting, which occurs when models learn to accurately anticipate results from training data but struggle to use new data, this is done. Even though a model may have operated flawlessly throughout development, it may diagnose real patients incorrectly. Furthermore, it can be challenging to comprehend what artificial intelligence models have learned, such as why they forecast a certain level of danger. This is a widespread issue with artificial intelligence systems generally, and it has given rise to an entire branch of study called explainable artificial intelligence.

The Black Dog Institute discovered a high likelihood of bias in 42 of the 56 studies it examined. In this case, a bias means that the model overestimates or underestimates the average suicide rate. Most of the data for the models came from electronic health records. However, some also contained information from clinical notes, self-report surveys, and interviews. The advantage of adopting artificial intelligence is that it can identify patterns missed by overburdened medical personnel and learn from massive amounts of data faster and more effectively than humans.

Although there has been improvement, the artificial intelligence approach to suicide prevention is not yet suitable for implementation in real-world applications. Suicide prediction is not the only strategy for lowering suicide rates and saving lives though. If an accurate prognosis doesn't result in a successful intervention, it is of no use.