AI helps identify psychiatric comorbidities in multiple sclerosis
Key takeaways:
- A machine learning model detected anxiety and depression using data from PHQ-9 and GAD-7 scores.
- The model was more successful in identifying depression than anxiety.
WEST PALM BEACH, Fla. — A machine learning model using existing clinical data may help clinicians screen for comorbid anxiety and depression in patients with multiple sclerosis, according to a poster presented at ACTRIMS.
“From what we know about depression and anxiety, in MS, it’s very common,” Braxton Phillips, MSc, a doctoral student from the Cumming School of Medicine at the University of Calgary and peer supporter for MS Canada, told Healio. “It’s an unmet need that clinicians don’t talk about as often as we should.”
Although common, screening for psychiatric comorbidities like anxiety and depression is difficult in a clinical setting. Phillips and colleagues investigated whether they could input clinical data that were already available into a machine learning model to “screen” for these conditions in persons with MS.
They collected baseline data on 944 individuals who enrolled in the Canadian Prospective Cohort Study to Understand Progression in Multiple Sclerosis (CanProCo), a 5-year undertaking of working-age adults aged 18 to 60 years with a mean time from MS diagnosis of 2.7 years, along with healthy controls. A total of 828 individuals (87.8%) had a diagnosis of either relapsing-remitting or primary progressive forms of MS.
The participants completed the Patient Health Questionnaire-9 (PHQ-9) and the Generalized Anxiety Disorder-7 (GAD-7), which offered a broad range of variables such as age, sex, diet, overall health, education, employment and additional comorbidities.
The machine learning model was built from a list of variables selected from the CanProCo dataset, then passed through a regression model (elastic net) that eliminates certain predictors to accurately predict scores of 10 and above on the PHQ-9 and GAD-7.
According to the results, the elastic net produced successful predictive models for both PHQ-9 and GAD-7 scores, with area under the curve values of 0.927 for the former and 0.813 for the latter.
The model also registered a sensitivity of 88.3% and specificity of 84.7% for PHQ-9 scores, as well as positive predictive (PPV) value of 70.7% for PHQ-9 scores of 10 and higher and negative predictive value (NPV) of 94.6% for scores lower than 10. These numbers with respect to the GAD-7 were 73.8%, 75.5%, 43.7% (PPV) and 91.8% (NPV), respectively.
Further, the researchers reported that the model identified higher average GAD-7 scores in patients with relapsing-remitting compared with primary-progressive forms of MS.
“With the information that we had, we did a pretty good job of predicting depression, especially,” Phillips noted. “A lot of this prediction came from their scores on a questionnaire that looks at their fatigue, which I think is interesting because there is a lot of overlap in depression and fatigue.”