Listening to the diagnosis of depression, deep learning into the field of mental disease diagnosis

Release date: 2016-07-14

Researchers at the University of Southern California have developed a new machine learning tool that can identify specific speech characteristics in patients with depression to aid diagnosis. The tool, called SimSensei, listens to the patient's voice during the face-up process. People with mental or neurological problems can reduce their vowel pronunciation, which may make the diagnostics unclear. This idea (of course) is not a substitute for the original human diagnosis, but it can add auxiliary, objective and weighty information to the diagnosis process.

Misdiagnosis of depression is a prominent problem in the medical field, especially for primary care doctors who often make such mistakes. In 2009, a meta-research covering 50,000 patients found that doctors identified only about 50% of the correct rate of depression, and the rate of false positives (diagnosing patients without depression) exceeded the rate of underreporting (will People with depression are diagnosed as normal, and the ratio is about 3 to 1. This is absolutely unacceptable.

But this is also understandable. Doctors (especially general practitioners) can over-diagnose the disease to a considerable extent for two reasons. First, it is almost safer to mistakenly diagnose a disease-free disease as sick without being diagnosed as disease-free. Second, every diagnosis faces various possibilities, and eliminating the uncertainty requires more expertise and more confidence.

One of the major problems in diagnosing depression is that depression is a heterogenous disease. It has a variety of causes and a variety of manifestations. A primary care physician may receive hundreds of patients each week, be exposed to various diseases, and they should summarize the results of psychiatric diagnosis from the many abnormal symptoms and interview-based observations reported by the patient. The challenge is conceivable. know. Therefore, tools like SimSensei have huge room for development. SimSensei tracks changes in speech related to depression and records them in detail. “Previous studies have revealed that people with depression often show dull or negative emotional reactions, no changes in tone, loudness and pitch are monotonous, language activity is reduced, speech rate is slowed, pause time increases, and pauses often change,” A related paper from the University of Southern California wrote, "In addition, the study found that the pronunciation in depression shows an increase in the stretch of the vocal cords and vocal cords."

This is obviously a problem for machine learning that makes predictions based on noise data. In general, speech analysis is one of the main concerns in this area.

The analysis done by this set of tools appears to be quite simple on the surface. It simplifies the patient's speech, retains only the vowels, and then analyzes the frequencies of the first and second formants (spectral peaks) of the vowels a, i, and u. The instruments involved in the first two parts of this analysis process are real speech detectors and associated formant trackers. The third part is the algorithm, which is actually a very long-lived machine learning method (produced in 1967), known as the k-means algorithm. The basic way of working is to grab the data sets and divide them into different clusters centered on a certain mean.

The result of the clustering is a triangle space/graph, at each inflection point the peaks of the vowels a, i and u. The area inside the triangle represents the vowel space, and this is what this algorithm calculates and presents. The presented space is then compared to the "standard" vowel space as a reference, and the measured indicators of depression (and post-traumatic stress disorder) are expressed proportionally.

“We measured the results of an automated evaluation of the vowel space in an experiment with 253 samples and found that for novel subjects reporting symptoms of depression and post-traumatic stress disorder, this novel method can detect their meta The voice space is significantly reduced.” The University of Southern California team concluded, “Our research demonstrates that this test is reliable when analyzing a portion of a total conversation or a limited amount of speech data, which means that the algorithm is practical. Finally, we successfully revealed that these tests have good statistical robustness in different demographic data and pronunciation speeds."

The results of the analysis show that there is not much difference in the vowel space ratio between patients with depression and non-depression, but the differences between them are sufficient to illustrate the problem. Perhaps the most significant problem in the study was the preliminary grading of depression and non-depression based on the self-reported assessment of the subjects. In addition, the reduction of vowel space may not be completely attributed to depression and post-traumatic stress disorder, and speech data under schizophrenia and Parkinson's disease conditions will be studied in the future.

Source: MotherBoard