A unique machine learning tool shows promise in predicting the onset of bipolar disorder (BD) years before it occurs, results of a new proof-of-concept study show.
Early results of an ongoing prospective cohort study show that factors such as suicide risk, generalized anxiety disorder, and parental physical abuse in individuals at age 18 years are predictive of a BD diagnosis at age 22.
Machine learning may be a “powerful tool” for improving the prognosis and early detection of BD at a time when the disease may be less refractory, the investigators note, adding that it may also predict treatment response.
This potential gain of 4 years in detection prior to disease onset could make a “huge difference” in the lives of young people, study author Francisco Diego Rabelo-da-Ponte, a PhD student at the Federal University of Rio Grande do Sul, Porto Alegre, Brazil, said in a release.
However, Rabelo-da-Ponte told Medscape Medical News that the study is a proof-of-concept study and that more research is needed.
The findings were presented at the 33rd European College of Neuropsychopharmacology (ECNP) Congress, which was held online this year because of the COVID-19 pandemic.
Long Diagnostic Delays
Although BD is the sixth leading cause of disability worldwide, identifying it is challenging, the researchers note. As a result, there is an average delay of 6 years between first symptoms and a clinical diagnosis. Furthermore, only 20% of individuals with BD who present with a depressive episode are diagnosed within the first year. Together, these factors have “harmful consequences,” the researchers say.
In an attempt to predict BD onset, the researchers conducted a prospective, population-based study. Known as the 1993 Pelotas Birth Cohort Study, it included data on 3778 individuals from 5249 participants who were enrolled as newborns.
Participants were assessed at birth and underwent assessments at age 11, 15, 18, and 22 years. The retention rate was 76.3%.
Participants completed the Mini–International Neuropsychiatric Interview at 22-year follow-up. At each of the preceding assessments, the investigators gathered data on demographic status, socioeconomic status, medical history, parental health, and other types of information.
The population was divided into a training dataset, which represented 80% of the cohort, and a testing dataset, which represented 20%. A machine learning protocol was used to correct for imbalances in the data, and an elastic net algorithm was employed to predict BD at age 22 years.
Of the participants, 255 (6.7%) received a BD diagnosis at their last visit. Of these, the majority had bipolar I disorder or BD not otherwise specified.
Increasing Accuracy
The algorithm predicted BD with increasing accuracy as follow-up progressed. The highest performance was at 18 years ― a full 4 years before BD diagnosis.
At this point, the algorithm had a balanced accuracy of 0.75, a sensitivity of 0.72, a specificity of 0.77, a positive predictive value of 0.18, and a negative predictive value of 0.97.
Using the 18-year follow-up data, the algorithm predicted BD at age 22 years with an area under the curve of 0.82 (95% CI, 0.75 – 0.88).
By far the greatest predictor of a subsequent BD diagnosis at 18 years was suicide risk, which was present in almost all patients. This was followed by generalized anxiety disorder, parental physical abuse, financial problems, and engaging in physical aggression.
However, the predictive positive value of the model was low, “which leads to a higher number of false positives,” the investigators write.
The results “suggest that it is possible to utilize psychosocial and clinical measures to predict bipolar disorder at the individual level,” write the researchers.
Rabelo-da-Ponte noted that the research team plans to assess the next wave of follow-up when participants reach 30 years of age.
He said that in the meantime, they would like to replicate their findings “in other birth cohorts, maybe the ALSPAC birth cohort in the UK or other birth cohorts in Brazil.”
No Shortcuts
Eduard Vieta, MD, PhD, chair of the Department of Psychiatry and Psychology at the University of Barcelona Hospital Clinic, Barcelona, Spain, noted in a press release that population cohorts “are extremely important to develop predictive models that may aid in the prevention of serious conditions” such as BD.
“What is most needed is replication and verification of the validity of the algorithm,” said Vieta, who was not involved in the research.
“The present study, hence, has its merits but is relatively small and needs replication in a separate, independent cohort; moreover, unusual findings such as the underrepresentation of bipolar II disorder need clarification as well,” he added.
Commenting on the findings for Medscape Medical News, Joseph F. Goldberg, MD, clinical professor of psychiatry at Icahn School of Medicine at Mount Sinai, New York City, said that BD “is an extremely hard entity to diagnose even by experts unless the observation point is florid hospitalized mania.”
Goldberg, who was not involved in the study, added that the differential diagnosis “is wide” and that many factors can influence its emergence and presentation.
He also underscored the low positive predictive value of the model and noted that the variables “have no specificity to any specific disorder.
“I would not encourage shortcuts to making accurate diagnoses or predicting future diagnoses without a very careful incorporation of the many other conditions that mimic bipolar disorder,” Goldberg said.
The 1993 Pelotas Birth Cohort Study was supported by the Wellcome Trust, the Brazilian National Research Council of the Brazilian Ministry of Health, and the National Institute for Science and Technology – Translational Medicine. The study authors, Vieta, and Goldberg have reported no relevant financial relationships.
33rd European College of Neuropsychopharmacology (ECNP) Congress: Abstract P250, presented September 15, 2020.
For more Medscape Psychiatry news, join us on Facebook and Twitter.
