Abstract

Background/Objectives: Bipolar disorder (BD) is one of the leading causes of disability worldwide, causing significant functional impairments in those affected. The heterogeneous course of BD renders the prediction of clinical progress and outcomes challenging, but it can be potentially enhanced with the use of artificial intelligence methods. In this systematic review, we aimed to examine the extant literature regarding the predictive accuracy of clinical functioning, illness affective state, relapse, and relevant predictors amongst patients with BD, using artificial intelligence methods. Methods: The study was guided by PRISMA and the Cochrane Handbook for Systematic Reviews. Six electronic databases were systematically searched from inception for relevant studies until July 2025 and relevant data were summarised in tables. The protocol of the review was registered on Prospero, ID: CRD42024590343. Results: Forty articles were included in this review. The area under the curve (AUC) values for clinical functioning, illness affective state, and relapse prediction were 0.59–0.72 (poor to acceptable), 0.57–0.97 (poor to outstanding), and 0.45–0.98 (poor to outstanding), respectively. Supervised, tree-based algorithms performed the best. Predictive factors included sociodemographic, clinical and psychological factors and wearable data, as well as speech and video recordings. Conclusions: Existing studies showed the potential of machine learning methods in the prediction of clinical progress and outcomes of BD (specifically functional status, affective state, and relapse) based on relevant collected variables. Longitudinal studies can further clarify and validate the associated predictive factors for earlier identification of those at risk of poorer prognosis to enhance management of BD.

1. Introduction

Worldwide, bipolar disorder (BD) affects 37 million people [1,2]. BD is a major contributor to global disability and during manic episodes; patients with BD can experience irritability, increased energy, decreased need of sleep, impulsivity, racing thoughts, and other symptoms. During depressive episodes, they can experience suicidality, depressed mood, loss of interest, guilt feelings, lack of energy and concentration, and changes in sleep and eating patterns. A substantial proportion of patients with BD experience disruptions in their psychological and social functioning as a result of their condition [3]. Furthermore, BD is associated with increased mortality as compared with the healthy population due to premature death from suicide and other comorbidities such as cardiovascular, cerebrovascular, and respiratory conditions [4].

BD is heterogeneous in its types, presentations, course, and response to different treatments, thus rendering the prediction of clinical progress and outcomes challenging. In this regard, the use of artificial intelligence methods such as machine learning algorithms have been used to evaluate specific variables, which allow for better prediction of the clinical progress and outcomes of BD. For example, studies had sought to predict clinical functioning and changes in clinical states and illness course using sociodemographic, clinical, and other data [5,6,7]. Specifically, Rabeloo-da-Ponte et al. [8] examined 267 participants (167 outpatients with BD and 100 unaffected volunteers) from Brazil and Spain and classified the cognitive functioning of their cohort based on years of education, number of hospitalisations, and age. In another study conducted in Korea [9], sleep data (duration of sleep, length of circadian rhythm, step counts) from wearables were found to predict depressive, mania, and hypomania mood states. A study conducted in Spain by Anmella et al. [10] that evaluated data from wearables found that motor activities (acceleration), poor sleep, and skin temperature were most predictive of relapses. Miola et al. [11] found in their supervised machine learning model that cardiovascular, neurological, osteo-muscular disorders, valproate use, and age were associated with medical admissions in patients with BD.

To date, systematic reviews on machine learning applications in BD have been largely limited to diagnostic accuracy focused on diagnosis [12,13,14,15,16], while others have examined monitoring approaches for relapse [17,18]. Only one review to date has explicitly examined the prediction of relapse, hospitalisation, and suicide, but this was restricted to studies using neuroimaging and clinical data [19]. This review extends the existing literature by synthesising evidence across a broader range of clinically relevant outcomes, including clinical functioning and affective state, in addition to relapse, and by examining a wider spectrum of predictive variables, including clinical, psychological, behavioural, and physiological markers. Hence, we conducted a systematic review to evaluate the predictive accuracy of machine learning models in predicting clinical functioning, illness affective state, and relapse, as well as to identify the key predictors associated with these outcomes in bipolar disorder.

2. Methods

The methodology of this review paper was guided by the Preferred Reporting Items for Systematic reviews and Meta-Analyses (PRISMA) [20] and the Cochrane Handbook for Systematic Reviews of Diagnostic Test Accuracy [21]. The protocol of the review was registered on Prospero, ID: CRD42024590343. Our review questions were as follows:What is the predictive accuracy for predicting clinical functioning, illness state, and relapse amongst patients with BD using machine learning methods?What are the relevant clinical variables identified in predicting clinical functioning, illness state, and relapse among patients with BD?

Predictive accuracy was determined by measures including area under the curve (AUC), accuracy, and balanced accuracy, wherever available. The initial search strategy was conducted using PubMed with the key words ‘(artificial intelligence)’ AND (‘bipolar’) AND (‘functioning’ OR ‘illness state’ OR ‘relapse’). This initial search identified all relevant keywords specific to the PICO model. The PICO model is a framework used to formulate focused research questions by defining the Population, Intervention, Comparison, and Outcome. The following keywords were used:

(1) The keyword for population was ‘bipolar disorder*’.

(2) The keywords for machine learning methodology were (Artificial intelligence) OR (machine learning) OR (natural language processing) OR (neural network) OR (data science) OR (deep learning) OR (digital biomarker*) OR (digital phenotyping).

(3) The outcomes were (a) function*, (b) hospitali* OR relapse, OR (c) (illness state) (Supplementary Table S1). Keywords were searched on six electronic databases, Scopus, ScienceDirect, PsycINFO, PubMed, Embase, and CINAHL, from database inception until July 2025.

2.1. Study Selection

The study selection process followed a rigorous, multi-stage approach. Two authors independently assessed the identified articles from the database search. After the removal of duplicate records, the remaining articles were screened based on titles and abstracts to ensure a broad initial capture. Then, full-text screening was done for all potentially relevant articles to confirm eligibility based on methodology and clinical focus. Studies were included if they (1) applied machine learning methods for predictive modelling; (2) focussed on patients with bipolar disorder; and (3) predicted clinically relevant outcomes such as functioning, illness affective states, and relapse. Only peer-reviewed articles printed in English were included. Studies were excluded if they (1) did not employ machine learning techniques; (2) did not focus on bipolar disorder or include outcomes of interest related to bipolar disorder; (3) were review articles, editorials, commentaries, or conference abstracts without full methodological descriptions; or (4) did not provide sufficient methodological information for evaluation. In addition, reference lists of included articles were also reviewed to identify any additional studies not captured in the initial database search.

2.2. Data Collection and Analysis

The data summarised were clinical setting factors, participants’ characteristics, specific clinical outcomes examined (functioning, illness affective state, relapse), variables examined, and main findings (see Table 1). Predictive accuracy summaries are found in Table 2.

Risk of Bias

Risk of bias was appraised using the Joanna Briggs Institute checklist for diagnostic test accuracy studies [22]. The Joanna Briggs Institute checklist for diagnostic test accuracy studies was selected because included studies of prediction modelling evaluated the performance of machine learning models in predicting clinical outcomes, which involved the assessment of model performance measures such as accuracy and related metrics. The other framework such as the Prediction model Risk of Bias Assessment Tool (PROBAST) was also considered when interpreting the findings. Two authors appraised the risk of bias of included studies. Discrepancies between raters were resolved through mutual discussion within the study team to reach consensus.

3. Results

Overall, a total of 1556 articles were initially identified, 114 full-text articles were evaluated, and 40 relevant studies were eventually included in this review (see Figure 1).

Regarding quality appraisal of the included articles, the Joanna Briggs Institute checklist for diagnostic test accuracy studies was used and three items (blinded interpretation of both the index test and reference standard results; appropriate interval between index test and reference standard) were excluded because they fall outside the scope of this review, focusing on clinical diagnostic accuracy rather than internal validation of artificial intelligence methods. Scores of zero to three are considered poor quality, while scores of four and above indicate moderate quality. The scores of the quality appraisal for the included papers (see Supplementary Table S2) ranged from three to seven, with most studies scoring about four to five (moderate quality).

The main details of the included studies are summarised in Table 1. In terms of the sites of the studies, the majority of studies (n = 26, 65%) were conducted in the West, namely (1) Americas (11 studies) such as the United States (n = 7) [23,24,25,26,27,28,29], Colombia (n = 1) [30], and Brazil (n = 3) [5,31,32], and (2) Europe (15 studies) such as Austria (n = 3) [33,34,35], Greece (n = 2) [36,37], Italy (n = 2) [11,38], Norway (n = 2) [39,40], Spain (n = 2) [10,41], Denmark (n = 1) [42], and Portugal (n = 1) [43]. One study recruited from more than one country, i.e., Brazil and Spain [8], and another study was based on a European project [44].

The rest of the included studies were from Asia (14 studies) such as Korea (n = 2) [9,45], China (n = 2) [7,46], Japan (n = 1) [47], and India (n = 1) [48]. Three studies recruited participants from a Turkish database [49,50,51], and three studies from the Audio Visual Emotion Challenge dataset [52,53,54]. One study recruited participants in Jordan (n = 1) (Ebrahim et al., 2018) and another from those who accessed a phone application (BiAffect) [6].

In terms of sample size, it ranged from six to 20,000 participants across the studies. The number of electronic health records and call entries ranged from 3085 to 77,296. Predictive items that were examined included sociodemographic and clinical variables (psychiatric, medical history factors) and data from pharmacological, biochemistry, MRI, social, psychological, physiological (sleep, activity), speech (acoustic, linguistic), and video (facial cues, eye gaze) domains.

3.1. Accuracy of Artificial Intelligence Methods in Predicting Clinical State or Outcomes

Predictive accuracies of studies are reported with accuracy, balanced accuracy, and/or area under the curve (AUC) (Table 2). AUC classifies true negatives and positives [56] with the following scorings: poor (0.50 to <0.70), acceptable (0.70 to <0.80), excellent (0.80 to <0.90), and outstanding (0.90 and above) [57]. Predictive accuracies are more specific for classification models in machine learning and include accuracy and balanced accuracy. Accuracy is the number of correct predictions divided by the number of total prediction number. Balanced accuracy is reported for imbalanced data and is the division of (sensitivity + specificity) by 2. The interpretation of scorings for accuracies are poor (<60%), acceptable (60% to <70%), good (70% to 90%), and very good (>90%) [58].

The included studies showed substantial heterogeneity in data sources, machine learning algorithms, outcome definitions, and predictors. Similarly, a wide variety of supervised algorithms was applied, and outcomes such as clinical functioning, affective state, and relapse were defined and operationalised differently across studies. As a result, while comparative summaries of predictive performance provide a useful overview of the field, differences across studies should be considered when interpreting variations in reported metrics.

Illness affective state prediction had the highest predictive accuracy, achieving an AUC of 0.57–0.97 (poor to outstanding) and accuracy of 45–100% (poor to very good), while illness relapse prediction had an AUC of 0.45–0.98 (poor to outstanding) and accuracy of 52–98% (poor to very good). Predictive accuracy for clinical functioning was the lowest, achieving an AUC of 0.59–0.72 (poor to acceptable) [31,32]. The AUC values in this review showed greater variability compared to those reported in another review examining relapse (AUC 0.71–0.98) and suicide (AUC 0.71–0.99) in bipolar disorder [19]. This likely reflects the greater heterogeneity of the included studies in the present review, whereas Amanollahi et al. [19] primarily included studies based on neuroimaging and clinical data that largely utilised internal validation. In addition, the present review includes a broader range of outcomes, such as clinical functioning and affective state, and captures emerging multimodal approaches, including wearable, behavioural, and speech data.

Among the studies that evaluated the performance of several algorithms, random forest consistently outperformed the other algorithms in five studies [27,30,31,34,55], achieving an AUC of 0.68–0.98 (poor to outstanding) and accuracy of 80–95% (good to very good). Another tree-based algorithm, gradient boosting, also performed well in two studies, as reflected by an AUC of 0.79–0.80 [7] and accuracy of 80–100% [53].

3.2. Functioning

Three studies assessed functioning as an outcome [8,31,32]. Sociodemographic and clinical data that were predictive of better functioning included higher education levels, fewer hospitalizations, and specific brain structures such as larger frontal cortex volume and smaller right lateral ventricle volume [32] Poorer functioning was associated with childhood abuse and neglect [31].

3.3. Illness Affective State Differentiation

Twenty studies evaluated variables associated with illness affective state. Two studies examined demographic and clinical variables [5] while one study evaluated data from electronic health records [7]. One study examined video recordings [49] and five studies examined speech recordings [23,24,38,50,51], whereas another five studies examined both video and speech recordings [6,48,52,53,54]. Four studies examined data captured by smartphone, including physiological [33,45] and phone activity [34,35]. Two studies evaluated physiological data from wearables [39], including a T-shirt wearable [44], while another study examined the keystroke data from BiAffect application users [55].

Predictors of Mood State

(i) Sociodemographic and clinical data

Relevant associated clinical variables included a history of hospitalisation, number of hospitalisations, and number of hypomanic, manic, or depressive episodes [5]. Conversely, the number of total episodes or number of mixed episodes did not differentiate mood state, nor did the number of suicidal attempts or co-morbidities such as substance dependence, anxiety, and eating disorders [5].

(ii) Video and speech recordings

For speech recordings, one study found that pause features (duration and number of pauses) could classify mood state better than duration and rate of speech [38]. Likewise, reduced pause ratio measure and greater vocalisation were found in hypomanic speech, while the reverse was true for depressed speech [24].

Several studies evaluate variables that can distinguish between mania, hypomania, and remission. A study using video recordings found that body posture and head movements were important variables [52]. The study also observed that minor, rather than large, movements were useful variables [52]. Conversely, another study found that integrated variables of various modalities (verbal, visual, and aural) performed the best [48]. Yet another study found that speech data was more important than video data [54]. The study also found that emotions in speech data could classify states in males better than in females [54].

(iii) Smartphone and wearables

In terms of data from wearables, voice, together with phone call data, performed better than phone call data alone [33,35]. However, activity data performed better than voice and phone data [33]. Notably, another study found that activity, together with voice and speech data, performed the best [35].

Mania state was observed to have unstable patterns of activity while euthymic state was associated with dramatic activity shifts [39]. Activity and light exposure were observed to be higher at night in the high-mood group, but higher during the day in the low-mood group [45]. Sleep irregularity was also prevalent in the high-mood group [45]. Among a group of BiAffect application users, variables predictive of mood variations were female, younger age, post-traumatic stress disorder, and attention deficit stress disorder [55].

3.4. Relapse

Seventeen studies evaluated relapse as an outcome. Three studies specifically examined depressive [27,43,47] and manic relapses [40,41,46], while four studies examined hospitalisation [11,28,29,30].

Seven studies examined sociodemographic and clinical data [11,25,27,28,29,30,43], one study examined MRI data [42], and two studies examined both sociodemographic and clinical data, as well as MRI data [26,41]. One study each examined video recordings [47] and speech recordings [46]. Six studies examined information from wearables including physiological data [9,10,36,37,40] and subjective mood [9].

Predictors of Relapse

(i) Sociodemographic and clinical data

Sociodemographic characteristics predictive of relapse included younger age [25] and both female [25] and male gender [28], in contrast to another study which did not find associations with gender, age, ethnicity, and years of education [43].

For clinical data, younger age of onset [41,43], longer duration of illness [43], frequent hospitalisations or emergency room visits [25,28], short hospitalisation stays [25,28], multiple comorbidities [28], disorders of cardiovascular, neurological and osteo-muscular systems [11], greater symptoms of borderline personality disorder [25], administration of valproate medication [30], and discharge to non-facility settings [28] were relevant predictors of relapse.

In terms of symptomatology, anhedonia [27], fatigue [27], impulsivity [41], and abnormal thought process [41] were associated predictors compared with low self-esteem, psychomotor retardation, poor concentration, sleeping problems, and poor appetite [27].

In terms of brain MRI data, increased amygdala activity [42] and reduction in inferior frontal white matter or cerebellum volume [41] were associated predictors of relapse. Severity of depressed mood was predicted by connectivity of supplementary motor area nodes and left dorsolateral prefrontal cortex with various cortical, motor, limbic, and cerebellar areas while severity of elevated mood was predicted by connectivity of right visual association region nodes and left fusiform with various insular, motor, limbic, and posterior cortices [26]. Patients with longer and more frequent mixed affective episodes had hypoactivity in emotional regulation areas [42].

(ii) Video recordings

In terms of video recordings, slower head movements were noted in people experiencing depression [47].

(iii) Wearables

Associated predictors from wearables included electrodermal activity (aggressiveness, anxiety) [10], acceleration (poor sleep, motor activities) [10,37,40], heart rate [40], skin temperature [10], and hand gestures [40]. More specifically, in depression, there was reduced sleep efficacy with greater mean steps at night but lower mean steps during the day, with greater circadian rhythm [9]. In mania, there was shorter sleep with similarly greater mean steps at night and circadian rhythm acrophase. Hypomania presented similarly to mania, except that it showed a shorter mean circadian rhythm phase [9].

4. Discussion

This review found that most of the included studies were from the West (65%) and mainly published in the last six years. Several key findings were noted. First, the range of predictive accuracies using machine learning models was relatively wide for functioning, affective state, and illness relapse. Second, the best performing method was supervised model tree-based algorithms, particularly random forest and gradient boosting. Third, heterogeneous predictive variables examined included data related to sociodemographic and clinical factors, pharmacological and MRI data, physiological measures, and speech and video recordings.

Overall, AUC values for prediction of affective state, illness relapse, and functioning were 0.57–0.97 (poor to outstanding), 0.45–0.98 (poor to outstanding), and 0.59–0.72 (poor to acceptable) respectively. The values were comparable with other studies using machine learning methods, albeit to diagnose BD (AUC 0.48–0.65) and schizophrenia (AUC 0.54–0.95).

The wide range of predictive performance observed across studies appears to reflect heterogeneity in outcome definitions, data sources, and modelling approaches, rather than random variation. Higher predictive accuracy was more frequently reported in studies using digital phenotyping or electronic health record data, particularly for affective state, relapse, and hospitalisation outcomes that rely on behavioural variables, while prediction of functional outcomes and single-modality models showed comparatively more modest or variable performance. Affective state prediction yielded the narrower range of predictive accuracies compared with functioning, as varying mood states may be associated with relevant clinical symptoms and physiological measures. Of note, prediction of relapse relies on the relevant combination of baseline measures for clinical progress over time. Prediction of functioning as an outcome variable was limited by the small number of studies to date and may not capture significant data points such as support system, personal coping, and intercurrent stressors within the machine learning studies. Likewise, several studies reported very high AUC values (0.97–0.98). These findings should be interpreted with caution due to limitations such as small datasets, lack of external validation, and potential overfitting. Consequently, such results might not be generalisable to external datasets or clinical settings.

This review found that tree-based algorithms under supervised learning models, particularly random forest and gradient boosting performed, the best. Similarly, another review that evaluated prognosis in psychosis found that random forest was amongst the best performing algorithms [59]. Tree-based models can handle complex relationships between variables, large numbers of predictors, and missing data [60], and hence are appropriate for clinical studies. In contrast, Cho et al. [45] found that Support Vector Machines outperformed other models in diagnosing mental disorders. In terms of the predictive accuracy in diagnosing BD specifically, a recent review found that deep learning, specifically the convolutional neural network, performed the best [61]. The variations in the performance of machine learning models across studies are dependent on the unique strengths and characteristics of each model. For instance, the number and type of predictive variables involved can influence the performance of the different models as Support Vector Machine algorithms work well with sparse data while gradient boosting machine and random forest algorithms perform well with significant data [45]. Other reviews that evaluated the prognostic outcomes of mental disorders did not comment on the best AI algorithm used [19,62,63]. As Amanollahi et al. noted, appropriate comparisons of the accuracy of AI algorithms need to take into account the specific outcomes and predictors examined across studies [19].

With regard to functioning, consistent with the current literature, higher educational level [64] and fewer hospitalisations [64,65] were predictive of better cognitive functioning, while childhood abuse and neglect were predictive of poorer functioning. Other previously reported positive predictors included preserved cognitive ability [64] and lithium therapy within first ten years of diagnosis [65] while negative predictors included substance use and presence of comorbidities [65], such as diagnosis of depression, mixed states, and presence of personality disorder [64]. Whilst a study in this review found that age of onset, illness duration, and family history of mood disorders were not discriminatory of cognitive functioning, these variables were associated with illness severity in BD in other studies [66,67].

In terms of association with affective states, this review found that pause features, vocalisation, and postures in speech and video recordings were associated with different illness states. Consistent with earlier studies, physiological data from wearables such as activity and sleep levels were useful variables to distinguish illness states [17]. Further work needs be carried out to evaluate the comparative predictive accuracy of individual versus combinatorial data stream in differentiating changing affective states.

In terms of relapse, findings from machine learning (ML) models both align with and diverge from the broader clinical literature. Sociodemographic variables such as age and gender demonstrated effects across studies. Younger age was identified as a relapse predictor, and is consistent with epidemiological data, which found that older age is associated with more favourable long-term outcomes, likely due to greater illness insight and treatment stability in older adults [68]. The mixed associations with gender mirror broader reviews, indicating that gender effects are modest and context-dependent, frequently mediated by clinical and psychosocial factors rather than acting as direct predictors of relapse itself [68].

Clinical and symptomatic predictors were more concordant with existing knowledge. Prodromal and behavioural changes, including mood lability, sleep disturbance, psychomotor shifts, anxiety, and impaired thought processes, have long been recognised as early warning signs of mood episode recurrence (e.g., elevated unusual thought content preceding relapse) in BD [69]. Beyond traditional clinical markers, machine learning models incorporating neurobiological data (e.g., altered amygdala activity, changes in white matter volumes, disrupted functional connectivity) highlight the potential of integrating objective biomarkers into relapse risk stratification. These neuroimaging findings are consistent with illness models positing that affective dysregulation and reward circuitry abnormalities underpin mood instability in BD [70].

Importantly, digital phenotyping signals derived from video and wearable sensors, such as reduced head movement in depressive states, electrodermal changes, heart rate, sleep patterns, activity acceleration, and circadian rhythm alterations, represent emerging objective predictors that complement subjective symptom monitoring. Some evidence suggests that physiological markers from wearables may detect relapse risk states that precede clinical presentation, aligning with research advocating for physiological monitoring to enhance predictive accuracy beyond self-reported symptoms [71]. Altogether, these findings suggest that multimodal predictive models—that is, combining sociodemographic, clinical, neurobiological, and digital phenotyping features—may offer the most promise for relapse prediction in BD, while acknowledging the need for larger, diverse datasets and external validation to improve generalisability and clinical utility.

There are several limitations of this review. First, the search strategy was limited to studies published in English, which may have resulted in the exclusion of relevant studies published in other languages. Second, this review paper may be vulnerable to publication bias as grey literature was excluded. Third, despite efforts to minimise selection bias through systematic screening of articles, some degree of subjectivity in the study selection process cannot be entirely excluded. Finally, there was considerable heterogeneity among the included studies in terms of methodology (e.g., datasets and sample sizes), study populations (e.g., clinical background), and outcomes examined. The small sample size in certain studies also limits the statistical strength of the findings. The studies also utilised different predictive variables, machine learning algorithms, performance metrics, and validation approaches, preventing quantitative synthesis such as meta-analysis. Consequently, the findings of this review were synthesised narratively. Future research could look into collaborative studies involving larger samples from multiple sites to enhance generalisability of research findings. Studies may also want to examine multimodal predictive models, that is, combining sociodemographic, clinical, neurobiological, and digital phenotyping features for the prediction of clinical outcomes in BD.

In conclusion, existing studies demonstrate the growing potential of machine learning methods to predict key clinical trajectories in BD, including functional outcomes, affective states, and relapse risk, using diverse clinical, biological, and digital data sources. However, current evidence remains heterogeneous and largely exploratory. Well-designed longitudinal studies with larger, more representative samples and external validation are needed to refine and confirm clinically meaningful predictors. Such advances could enable earlier identification of individuals at heightened risk of poor prognosis, support more timely and targeted interventions, and inform personalised management strategies to improve long-term outcomes in BD.

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Supplementary Materials

The following supporting information can be downloaded at: https://www.mdpi.com/article/10.3390/brainsci16040415/s1, Table S1: Search keywords; Table S2: Quality Appraisal.