Source-linked AI summary
A signature-based machine learning model for bipolar disorder and borderline personality disorder
Imanol Perez Arribas, Kate Saunders, Guy Goodwin, Terry Lyons
TL;DR
Mobile mood monitoring offers more precise psychiatric data, but analyzing oscillatory mood reports remains challenging. This paper uses signature-based machine learning to classify diagnostic groups and predict subsequent mood, reporting strong classification and next-mood prediction results.
Problem
Psychiatric diagnoses often rely on retrospective mood recall, while oscillatory symptom time series are challenging to analyze.
Source: §1 · p. 2Method
A signature-based machine-learning model classifies participants from mood reports and predicts next mood from the last 20 observations.
Source: §1 · p. 2, §2.3 · p. 6Results
75% of participants were correctly classified versus 54% with the mean-score model; next-mood accuracy was 89–98% in healthy, 82–90% in bipolar, and 70–78% in borderline participants.
Source: §3.1 · p. 6, §3.2 · p. 8Takeaways & Limitations
The authors conclude that the model was effective on 20-observation mood streams and may be robust and applicable beyond the study cohort.
Source: §4 · p. 9
Abstract
from arXiv · showhide
Mobile technologies offer opportunities for higher resolution monitoring of health conditions. This opportunity seems of particular promise in psychiatry where diagnoses often rely on retrospective and subjective recall of mood states. However, getting actionable information from these rather complex time series is challenging, and at present the implications for clinical care are largely hypothetical. This research demonstrates that, with well chosen cohorts (of bipolar disorder, borderline personality disorder, and control) and modern methods, it is possible to objectively learn to identify distinctive behaviour over short periods (20 reports) that effectively separate the cohorts. Participants with bipolar disorder or borderline personality disorder and healthy volunteers completed daily mood ratings using a bespoke smartphone app for up to a year. A signature-based machine learning model was used to classify participants on the basis of the interrelationship between the different mood items assessed and to predict subsequent mood. The signature methodology was significantly superior to earlier statistical approaches applied to this data in distinguishing the participant three groups, clearly placing 75% into their original groups on the basis of their reports. Subsequent mood ratings were correctly predicted with greater than 70% accuracy in all groups. Prediction of mood was most accurate in healthy volunteers (89-98%) compared to bipolar disorder (82-90%) and borderline personality disorder (70-78%).
1 Introduction
Mobile mood monitoring may improve psychiatric phenotyping, but oscillatory symptom data are difficult to analyze. The paper applies signature-based machine learning to classify diagnostic groups and predict subsequent mood.
1 Introduction
- Mobile monitoring can capture more precise measures of psychopathology, while oscillatory symptom patterns pose an analytic challenge.The introduction contrasts these measures with diagnoses reliant on retrospective recall.Source: §1 · p. 2
- A signature-based model classifies participants from mood reports and predicts their subsequent mood, treating classification and forecasting in a similar way.The study uses daily mood data from participants with bipolar disorder, borderline personality disorder, and healthy volunteers.Source: §1 · p. 2
2 Methods
Participants’ daily reports were organized as short, normalized streams and represented as paths for signature-based analysis. The methods applied that representation to diagnostic-group classification and next-mood prediction.
2.1 Data
- 130 participants recorded six daily mood categories on a 7-point scale using a smartphone app; reports were grouped into 20-observation streams and normalized.The cohort included 48 participants with bipolar disorder, 31 with borderline personality disorder, and 51 healthy participants; compliance was 81.2%.Source: §2.1 · p. 3
2.2 Group classification
- For group classification, each 20-observation stream was represented as a 7-dimensional path and transformed into a truncated signature, with orders 2, 3, and 4 considered.The comparison used mean mood scores by category; classification performance was assessed with accuracy, sensitivity, specificity, PPV, and ROC AUC, with bootstrapping at order 2.Source: §2.1 · p. 3, §2.2 · pp. 4–5
2.3 Predicting the mood of a participant
- For next-mood prediction, regression used the last 20 observations to predict the next report, with separate models trained for each clinical group.Predictions were evaluated using MAE and the percentage within one point of the observed category score.Source: §2.3 · p. 6
3 Results
The signature-based model classified participants into their diagnostic groups and predicted subsequent mood scores. Classification exceeded the mean-score baseline, while mood prediction was more accurate in healthy participants than in either clinical group.
3.1 Predicting which clinical group a participant belongs to
- 75% of participants were classified correctly, significantly above the naive mean-score baseline’s 54%; the model performed best at order 2.Across 100 bootstrap samples, mean accuracy was 74.85% (SD 2.05), suggesting stable results.Source: §3.1 · p. 6, Table 1 · p. 7, Figure 3 · p. 8
- Healthy participants had fewer incorrect classifications than either clinical group, and the model clearly distinguished healthy from clinical participants.Source: §3.1 · p. 6, Table 2 · p. 7
- Participant-level periods characteristic of each group clustered in the correct triangle corners, more consistently for healthy than clinical participants.Source: §3.1 · p. 8
3.2 Predicting the mood of a participant
- Next mood scores were correctly predicted 89-98% of the time in healthy participants, 82-90% in bipolar participants, and 70-78% in borderline participants.Source: §3.2 · p. 8, Table 5 · p. 9
- Against last-day-score prediction, the model clearly outperformed benchmark accuracies of 61-92% for healthy, 46-67% for bipolar, and 44-62% for borderline participants.Source: §3.2 · p. 8
- Bootstrap histogram concentration suggested stability, while Figure 4 highlighted diagnostic-group separation; classification averaged 74.85% correct (SD 2.05).Source: §3.1 · p. 6, Figure 3 · p. 8, §3.2 · p. 9
- The model correctly predicted mood even in highly oscillatory anxiety paths from participants with bipolar disorder.Source: §3.2 · p. 9
4 Conclusion
The signature-based model was effective on 20-observation mood streams, and its construction suggests robustness beyond the study cohort.
- The signature-based model effectively predicted mood from participants’ streams of 20 observations.Source: §4 · p. 9
- The authors present the signature-based model as a novel approach to studying psychiatric data.Source: §4 · p. 9
- Because the model was built without being shaped by the data’s nature or structure, the authors suggest it may apply beyond the study cohort.Source: §4 · p. 9