Individualized identification of euthymic bipolar disorder using the Cambridge Neuropsychological Test Automated Battery (CANTAB) and machine learning

Abstract

Background: Previous studies have reported that patients with bipolar disorder (BD) present with cognitive impairments during mood episodes as well as euthymic phase. However, it is still unknown whether reported neurocognitive abnormalities can objectively identify individual BD patients from healthy controls (HC).

Methods: A total of 21 euthymic BD patients and 21 demographically matched HC were included in the current study. Participants performed the computerized Cambridge Neurocognitive Test Automated Battery (CANTAB) to assess cognitive performance. The least absolute shrinkage selection operator (LASSO) machine learning algorithm was implemented to identify neurocognitive signatures to distinguish individual BD patients from HC.

Results: The LASSO machine learning algorithm identified individual BD patients from HC with an accuracy of 71%, area under receiver operating characteristic curve of 0.7143 and significant at p=0.0053. The LASSO algorithm assigned individual subjects with a probability score (0-healthy, 1-patient). Patients with rapid cycling (RC) were assigned increased probability scores as compared to patients without RC. A multivariate pattern of neurocognitive abnormalities comprising of affective Go/No-go and the Cambridge gambling task was relevant in distinguishing individual patients from HC.

Limitations: Our study sample was small as we only considered euthymic BD patients and demographically matched HC.

Conclusion: Neurocognitive abnormalities can distinguish individual euthymic BD patients from HC with relatively high accuracy. In addition, patients with RC had more cognitive impairments compared to patients without RC. The predictive neurocognitive signature identified in the current study can potentially be used to provide individualized clinical inferences on BD patients.

Keywords: Bipolar disorder; CANTAB; Machine learning; Neurocognition; Rapid cycling.

Figure 1

A flow diagram showing the LASSO algorithm training and testing process using a leave-one-out cross-validation (LOOCV) approach.

Figure 2

A ‘Confusion matrix’ and a receiver operating characteristic (ROC) curve used to calculate LASSO algorithm's prediction accuracy, sensitivity, specificity and area under receive operating characteristic curve.

Figure 3

A) A bar graph showing model coefficients assigned to the ‘most relevant’ CANTAB neurocognitive variables by the LASSO algorithm. These neurocognitive variables were the most relevant in distinguishing individual euthymic BD patients from HC. Positive model coefficients reflect higher scores in BD patients as compared to HC, whilst on the contrary negative coefficients represent lower scores BD patients. B) The number of commission errors in response to negative stimuli was among the ‘most relevant’ variable predicting membership to BD group. Compared to HC patients with BD committed a greater number of commission errors when exposed to negative stimuli which may indicate the presence of the affective cognitive bias commonly observed in BD. Predictor variable coefficients should be interpreted with caution as they are selected to maximize prediction accuracy as opposed to represent individual coefficient accuracy. Error bars represent the standard error of the mean - estimated by the standard deviation divided by the square root of the sample size.

Figure 4

A) A bar graph comparing predicted probability scores between HC, BD without RC and BD with RC. Individual probability scores were ‘normalized’ by subtracting (0.5) for visualization purposes only and therefore HC had negative probability scores while BD patients had positive probabilities. Analysis of non-parametric Kruskal-Wallis statistical test comparing predicted probability scores between healthy controls, BD patients without RC and BD patients with RC was performed. Predicted probability scores in BD patients with RC and healthy control differed significantly (Kruskal-Wallis test p = 0.015) whilst BD without RC were less distinguishable from HC (Kruskal-Wallis test p = 0.073). B) A bootstrapping technique was used to estimate the distribution of the mean predicted probability for HC, BD patients without RC and BD patients with RC. For each group, the predicted probabilities scores were resampled with replication and the corresponding mean was calculated 100000, which led to a smooth estimated density function of the predicted probabilities.