Pediatric Bipolar Disorder (BD) is estimated to affect about 2% of youth and often presents with subsyndromal symptoms of mood dysregulation during childhood that eventually develops into a clinical picture meeting full criteria for bipolar disorder. The identification of children with emerging bipolar disorder is extremely challenging in clinical practice. Symptoms frequently associated with bipolar disorder, such as increased activity and impulsivity, are also characteristic of Attention Deficit Hyperactivity Disorder (ADHD); thus, many children with early BD are identified as having ADHD prior to a formal diagnosis of BD. Similarly, some children with BD present with irritability, sadness or anxiety and may be treated for major depressive disorder or an anxiety disorder prior to having manic or hypomanic symptoms.
While correlational studies have suggested that bipolar disorder is more likely to develop in children with early onset and greater severity of mood symptoms, family history of BP disorder, and severe emotional dysregulation (Uchida et al, 2015), it has been difficult to use this information in a clinical setting to reliably predict which children will develop bipolar disorder in the future. In a recent study, Mai Uchida, MD and her colleagues in the Clinical and Research Programs in Pediatric Psychopharmacology and Adult ADHD at Mass General have taken advantage of data collected from a large cohort of children followed over ten years to determine whether machine learning could be used to reliably predict the future development of BP-I disorder.
The analysis included male and female children who had participated in one of two longitudinal case-control family studies and were between the ages of 6 and 18 at study entry. Participants were assessed at baseline and over the course of 10 years of follow-up. In addition to providing sociodemographic data, children were assessed using psychometric scales, structured diagnostic interviews, and cognitive and social functioning assessments. Children meeting criteria for BD-I at baseline and their siblings, as well as children with ADHD at baseline, were excluded. Data from a total of 492 children (52% male, mean age at entry 11.1 years) were used to generate the predictive model.
In this cohort, 45 children (10%) developed bipolar disorder over the course of the ten year follow-up period. The machine learning algorithm generated a model that could accurately predict the development of BP-I disorder with 75% sensitivity, 76% specificity, and an Area Under the Receiver Operating Characteristic Curve of 75%. (ROC-AUC combines sensitivity and specificity and is a measure of the overall performance of a diagnostic test.) The model had a false positive rate of 21.6% and a false negative rate of 3.1%.
The researchers also identified which factors measured at baseline carried the most weight in terms of the differentiating children and adolescents who developed BP-I disorder from those who did not. All of the top seven features were drawn from the Childhood Behavior Checklist (CBCL): CBCL total t-score, CBCL Externalizing t-score, CBCL AAA (Aggression/Anxiety-Depression/Attention) profile t-score, CBCL Internalizing t-score, CBCL School Competence t-score, CBCL Anxious/Depressed t-score and CBCL Aggressive t-score.
What’s Next?
This study provides the first quantitative model for the prediction of bipolar disorder, type I in children. Being able to reliably predict which children are most likely to develop bipolar disorder would help clinicians to assess children with emergent psychopathology and to choose appropriate treatments. Such a model could alert clinicians and caregivers to improve monitoring over a 10-year period. This is especially important since frequently children with BD present with depression or ADHD symptoms, and the medications used to treat these disorders may actually worsen symptoms in a child with bipolar disorder.
This sort or predictive model may also give us the opportunity to design and test interventions that could decrease the likelihood of developing bipolar disorder or mitigate the severity of its symptoms and impact on functioning. Such machine learning algorithms must first be tested in other populations. The authors note that one of the limitations of this study is that the population used to test the algorithm was clinically referred and mostly Caucasian.
One of the most exciting things about this study is the relative ease and cost-effectiveness of gathering clinical information to predict future risk. The most robust predictors of risk are all elements of the CBCL. This questionnaire is completed by a parent or caregiver, and CBCL scores are generated using computer software; it would require no additional time or effort from the clinician.
The CBCL is now used widely in many pediatric offices as part of well-child visits. While typically used to screen for syndromal mental health problems, this type of screening could be used to identify children at risk for bipolar disorder in settings that are not equipped with psychiatric expertise. Those children identified with subsyndromal symptoms could be monitored more closely or may be eligible for interventions that reduce risk for developing syndromal psychiatric symptoms.
Read More
Uchida M, Bukhari Q, DiSalvo M, Green A, Serra G, Hutt Vater C, Ghosh SS, Faraone SV, Gabrieli JDE, Biederman J. Can machine learning identify childhood characteristics that predict future development of bipolar disorder a decade later? J Psychiatr Res. 2022 Dec;156:261-267.
Uchida M, Serra G, Zayas L, Kenworthy T, Faraone SV, Biederman J. Can unipolar and bipolar pediatric major depression be differentiated from each other? A systematic review of cross-sectional studies examining differences in unipolar and bipolar depression. J Affect Disord. 2015 May 1;176:1-7
