This is the first quantitative approach to predict bipolar disorder, offering sensitivity and specificity of 75% and 76%, respectively, reported lead author Mai Uchida, MD, director of the pediatric depression program at Massachusetts General Hospital and assistant professor of psychiatry at Harvard Medical School, Boston, and colleagues. With further development, the model could be used to identify at-risk children via electronic medical records, enabling earlier monitoring and intervention.
Dr. Mai Uchida
“Although longitudinal studies have found the prognosis of early-onset mood disorders to be unfavorable, research has also shown there are effective treatments and therapies that could significantly alleviate the patients’ and their families’ struggles from the diagnoses,” the investigators wrote in the Journal of Psychiatric Research. “Thus, early identification of the risks and interventions for early symptoms of pediatric mood disorders is crucial.”
To this end, Dr. Uchida and colleagues teamed up with the Gabrieli Lab at MIT, who have published extensively in the realm of neurodevelopment. They sourced data from 492 children, 6-18 years at baseline, who were involved in two longitudinal case-control family studies focused on ADHD. Inputs included psychometric scales, structured diagnostic interviews, social and cognitive functioning assessments, and sociodemographic data.
At 10-year follow-up, 10% of these children had developed bipolar disorder, a notably higher rate than the 3%-4% prevalence in the general population.
“This is a population that’s overrepresented,” Dr. Uchida said in an interview.
She offered two primary reasons for this: First, the families involved in the study were probably willing to be followed for 10 years because they had ongoing concerns about their child’s mental health. Second, the studies enrolled children diagnosed with ADHD, a condition associated with increased risk of bipolar disorder.
Using machine learning algorithms that processed the baseline data while accounting for the skewed distribution, the investigators were able to predict which of the children in the population would go on to develop bipolar disorder. The final model offered a sensitivity of 75%, a specificity of 76%, and an area under the receiver operating characteristic curve of 75%.
“To the best of our knowledge, this represents the first study using machine-learning algorithms for this purpose in pediatric psychiatry,” the investigators wrote.
Integrating models into electronic medical records
In the future, this model, or one like it, could be incorporated into software that automatically analyzes electronic medical records and notifies physicians about high-risk patients, Dr. Uchida predicted.
“Not all patients would connect to intervention,” she said. “Maybe it just means that you invite them in for a visit, or you observe them a little bit more carefully. I think that’s where we are hoping that machine learning and medical practice will go.”
When asked about the potential bias posed by psychiatric evaluation, compared with something like blood work results, Dr. Uchida suggested that this subjectivity can be overcome.
“I’m not entirely bothered by that,” she said, offering a list of objective data points that could be harvested from records, such as number of referrals, medications, and hospitalizations. Narrative text in medical records could also be analyzed, she said, potentially detecting key words that are more often associated with high-risk patients.
“Risk prediction is never going to be 100% accurate,” Dr. Uchida said. “But I do think that there will be things [in electronic medical records] that could guide how worried we should be, or how quickly we should intervene.”