AI is offering new ways to monitor mental health and depression. Patients must consult a psychiatrist directly in order to receive a diagnosis of mental health issues using traditional procedures. This takes a lot of time and energy.Â
Mental health practitioners have previously employed AI. For instance, large language models (LLMs) may search through transcribed interviews for contextual clues and speech patterns indicative of psychiatric illnesses. However, text-based AI has drawbacks. Language barrier, cultural quirks, and varying proficiency levels can distort outcomes. Additionally, language biases in society have been demonstrated to be reflected in LLMs; in one research, an LLM was unable to detect depression in black patients when compared to their white counterparts. In many cases the AI text-based diagnoses ended up with hallucinations. It means the results look interesting, that it is basically random and do not reflect real mental well-being.
For this reason, the new techniques being developed focus on the way words are delivered rather than on specific phrases. For instance, an AI model created by scientists at China’s South-Central Minzu University searches for minute variations in a patient’s speech. According to the researchers’ premise, those who are depressed could have unique speech patterns that are too faint for the human ear to pick up on.Â
To assist the model recognize complicated audio patterns, the system employs a process called “pre-training,” in which the model is initially exposed to vast volumes of ordinary speech. These patterns might include changes in voice quality, pitch fluctuation, and rhythm that are normally undetectable to the human ear. Without having to comprehend the words themselves, this pre-training serves as a linguistic tuning fork, enabling the system to detect complex speech fluctuations that could indicate sadness. With the aid of recordings of depressed patients, the researchers subsequently modified, or “fine-tuned,” this all-purpose system especially for depression identification.
The precision of this refined approach was astounding. The method was 96% successful in detecting the presence of depression in a binary classification task and 95% accurate when asked to classify its severity into four levels (no depression, mild, moderate, and severe) using a single clinical rating scale, according to findings published in June in Nature Scientific Reports.
Other approaches are also working. Researchers at Paris’s Sorbonne University have created a technique that uses sound waves captured by a smartphone app to identify a number of mental health issues. Initially, the sound waves are transformed into spectrograms, which are visual maps that show the frequency and loudness changes of a voice over time. The model then looks for characteristics of several mental diseases, such as anxiety, sadness, sleeplessness, and exhaustion, in each unique spectrogram.
Source: Economist
