ABSTRACT
Background
There is broad agreement that the onset of bipolar disorders (BD) can be predicted by using combined estimates of familial, genetic and clinical risk. However, there is a lack of consensus about the operationalisation of different risk attributes (e.g., symptoms vs. sub-threshold syndromes; disorder-specific polygenic risk scores [PRS] vs. multiple-disorder PRS dimensions) and their utility for predicting bipolar type 1 (BD-I) and 2 (BD-II). Likewise, analyses often fail to consider the optimal model for predicting outcomes where true cases will be in the minority.
Methodology
Proof of concept study employing an ensemble machine learning approach (Boosting) to develop models for classifying BD cases vs. non-cases using different combinations of risk attributes extracted from a database from a prospective longitudinal follow-up of twin and non-twin siblings in the peak age range for onset of major mental disorders.
Results
Of 1473 participants (mean age 26.3; female = 866), 104 developed BD-I (n = 30) or BD-II (n = 74). The best performing Boosting classification had an overall area under the receiver operating curve (AUROC) of 85.1% (95% Confidence Intervals: 80%, 88%); correctly identifying 86.7% BD cases. Variables with greatest relative influence were, in rank order: depressive symptoms, psychotic symptoms, a BD-Schizophrenia PRS dimension, hypomanic symptoms, and family history of BD. The model accurately classified 89% of manic cases but only 68% of hypomanic cases.
Conclusions
Improving the accurate prediction of BD onset would benefit from greater consensus regarding the operationalisation of known risk attributes and selection of analytic models that consider sample imbalances.