Based on retrospective genetic and health registry data, a multidiagnostic deep-learning computer model very precisely determined the mental disorder diagnoses of major depressive disorder, schizophrenia spectrum disorders, attention-deficit/hyperactivity disorder (ADHD), autism spectrum disorder and bipolar disorder when resembling a clinical setting prior to first examination. A researcher says that the model should be developed further into a prospective tool to help doctors.
Correctly diagnosing mental disorders as early as possible is important, since this enables people to be treated rapidly and reduces the negative effects.
However, early diagnosis is not that easy. One reason is the lack of objective markers in determining mental disorder diagnoses such as schizophrenia, ADHD, autism, bipolar disorder and major depressive disorder.
A new study shows that an advanced computer model made very precise retrospective diagnoses based on existing and available genetic and health registry data.
In the long term, the goal is that the computer model can assist general practitioners prospectively in the early assessment of patients so that they can be referred to specialists for possible treatment.
“We want to create a clinically useful support tool that uses existing data and can contribute to early diagnosis of people with a mental disorder. The earlier we can make the correct diagnosis, the faster we can get people into relevant treatment and the less burdensome the mental disorder can end up being,” explains a researcher behind the study, Michael Eriksen Benros, Clinical Professor and Consultant, Mental Health Centre Copenhagen and Department of Clinical Medicine, University of Copenhagen.
The research has been published in JAMA Psychiatry.
Most of the study was carried out through collaboration between the group of Michael Eriksen Benros, Professor, Mental Health Center Copenhagen, where Rosa Lundbye Allesøe was a PhD Fellow on the project, and with Simon Rasmussen, Associate Professor at the University of Copenhagen.
Model based on comprehensive registry data
The researchers trained a deep-learning algorithm to search retrospective patient data for events that would indicate a possible diagnosed mental disorder in their future.
To develop the model, the researchers used registry and genetic data for 63,535 individuals diagnosed and not diagnosed with a mental disorder. The data are from the iPSYCH database, with genetic data from blood samples taken at birth, the National Patient Registry, the Birth Registry, the CPR Registry and several other national registries in Denmark.
The model thus had access to information about each individual’s genetic risk of developing various mental disorders and to other types of health data that can indicate a possible later risk of developing a diagnosed mental disorder such as low birth weight, premature birth or a family history of mental disorder.
“Many factors can contribute to developing mental disorders, and doctors cannot readily keep track of them all. A prospective decision support tool that can collect the relevant information and calculate a risk score may therefore be relevant to support the doctor in the diagnostic process, similar to when doctors use X-rays to aid in determining whether a foot is broken,” says Michael Eriksen Benros.
Model is very accurate
The testing of the model shows that it can determine quite accurately the later mental disorder.
The researchers gave the model access to patient data that was updated until a possible diagnosed mental disorder, and it could then determine who developed such a disorder and who did not.
The accuracy of a predictive tool is measured by the area under the curve (AUC), where 1.0 indicates 100% accuracy.
The computer model had an AUC of 0.84 for schizophrenia spectrum disorders; 0.79 for bipolar disorder; 0.77 for autism spectrum disorder; and 0.74 for both ADHD and major depressive disorder.
The model also had an AUC of 0.72 in determining the severity of the disorders, although the estimate was based only on pre-diagnosis information, and if more information is included after the hospital contact leading to the diagnosis, the prediction of the severity of the disease course could likely be further improved.
“I was surprised that the model is already so good at especially assessing the risk of schizophrenia spectrum disorders. However, there are many opportunities to improve the model further by adding clinical data at the time a person is prospectively examined,” explains Michael Eriksen Benros.
Model can be improved further
Michael Eriksen Benros says that further developing the model will increase the model’s accuracy even more and then it can be tested in clinical studies that can pave the way for implementing a model 2.0 version.
Among other enhancements, the researchers want to include data from electronic patient records, brain scans and blood sampling in the model.
This will make it even better and will also make the integration of the model into doctors’ work tools easier, so that it can smoothly support them in the early assessment of patients with suspected mental disorders.
“Clinical psychiatry has no such models yet, but we want to develop them. In this respect, we also need to determine how to integrate this and meet the needs of clinicians and patients,” says Michael Eriksen Benros.
He elaborates that the model can also highlight the factors on which it bases its predictions, which can help doctors to focus on relevant further examinations of the patient to provide the most optimal treatment and prevention efforts.
“The model may identify some factors that provide a basis for suspecting a mental disorder, which doctors can then supplement with their own investigations. Moreover, this may also benefit health workers more broadly and strengthen the case for referring a patient for further specialised examination,” concludes Michael Eriksen Benros.
