Prospects of machine learning applications in affective disorders

Full Text

INTRODUCTION

Mental disorders are among the most significant medical and social challenges for Russian and global healthcare systems. In Russia, approximately 40% of the population exhibit impaired mental health [1], whereas approximately one in eight individuals worldwide, an estimated 970 million people, live with a mental disorder [2]. Among them, approximately 280 million are diagnosed with depression and approximately 40 million with bipolar disorder (BD).¹ Psychiatry is one of the few fields in medicine where objective biological assessment methods remain largely unavailable [3], hindering timely and accurate diagnosis. Moreover, treatment remains ineffective in many cases [4]. For instance, a correct diagnosis of BD typically takes 5–10 years [5], and only one-third of patients with depression achieve remission following first-line therapy [6, 7].

Furthermore, the use of new technologies is expanding across different medical fields to improve diagnosis and treatment [8]. The rise of telemedicine, accelerated by the COVID-19 pandemic,² has enabled the continuous collection of text, audio, and video data, which was previously limited to controlled clinical trials. Accumulated large, heterogeneous datasets, commonly referred to as big data, cannot be processed using traditional statistical approaches [9]. Recent advances in natural language processing, speech recognition, and video analysis have increased interest in leveraging advanced computational methods, such as machine learning and deep learning [10]. Over the past decade, the number of studies found in PubMed with the keywords artificial intelligence, machine learning, psychiatry, and mental health has increased approximately 50-fold, now exceeding 2000. These novel data processing techniques may enable integrating multimodal information to enhance risk assessment, diagnosis, treatment selection, and relapse prediction in mental disorders. Additionally, they allow identifying new relationships between data points, supporting more nuanced and differentiated diagnosis and classification approaches [3, 10, 11]. Artificial intelligence (AI) technologies facilitate personalized diagnostic and therapeutic strategies tailored to individual patient characteristics, which is a central goal of precision psychiatry [12]. Clinical psychiatrists are increasingly interested in AI application, hoping it will enable more accurate diagnosis, optimize treatment selection, and ease administrative burdens [13]. However, concerns have emerged regarding potential job displacement if AI systems become capable of replacing human clinicians [14]. Despite the increasing research on AI applications in psychiatry, the feasibility of widespread clinical implementation remains uncertain. Moreover, the scalability and utility of these technologies in real-world settings remain debatable [15, 16].

Given the growing interest in digital tools and the need for novel innovative approaches to psychiatric care, this study explored current and emerging applications of AI in clinical practice using depression and BD as illustrative examples. Furthermore, the potential benefits and limitations of the included studies were assessed.

SEARCH METHODOLOGY

The search of articles was conducted from January to February 2024 in the databases PubMed, Google Scholar, and eLibrary using the following keywords and their combinations: психиатрия/psychiatry, психическое здоровье / mental health, психическое расстройство / psychiatric disorder, депрессия/depression, депрессивный эпизод / depressive episode, рекуррентное депрессивное расстройство (recurrent brief depression), биполярное расстройство / bipolar disorder, major depressive disorder, машинное обучение / machine learning, глубокое обучение / deep learning, and искусственный интеллект / artificial intelligence. Additional sources were identified by screening reference lists of relevant publications, focusing on studies that examined the application of each data type in patients with depression and BD.

This review included Russian- and English-language publications over the past 10 years, comprising systematic and narrative reviews, meta-analyses, and original studies on the use of AI technologies in diagnosing depression and BD. Reviews addressing the challenges of implementing AI in psychiatric practice were also included. The last search queries were performed on February 15, 2024. Studies focusing on AI applications in children and adolescents (<18 years) and publications without accessible full texts were excluded.

The final sample comprised 114 publications: 72 original study articles, 35 narrative reviews, and 7 systematic reviews.

ARTIFICIAL INTELLIGENCE, MACHINE LEARNING, AND DEEP LEARNING

AI refers to algorithms capable of reasoning, learning, planning, and performing actions that mimic human cognitive functions. This umbrella term encompasses machine learning (ML) and deep learning (DL) [17, 18]. ML is a subset of AI involving models that “learn” by identifying patterns in provided datasets and applying such learning to predict properties of new data.

ML can be categorized as supervised, semi-supervised, or unsupervised. Table 1 summarizes the specific methods used in each learning paradigm. Supervised learning implies the presence of labeled data, wherein model training directly depends on these labels [19, 20]. This approach is the most commonly used in medical research and includes classification and regression techniques [19, 21, 22]. Supervised classification is considered more suitable for diagnostic and prognostic tasks, predicting the presence or absence of mental disorders and course of illness [23] and treatment response and relapse risk. In turn, regression models may be employed to predict individual symptom severity during the disease course [24] and their changes over time following treatment [25].

 

Table 1. Categories of machine learning, operational principles, methods, and potential applications [10, 41]

Learning Cate-gory	Operational Principle	Methods	Applications
Supervised Learning	Data are provided with labels; the model learns to associate input with outcomes	• Classification methods: – Naive Bayes classifier – Support vector machines – Random forest – Adaptive boosting (AdaBoost) • Regression methods: – Linear regression • k-nearest neighbors	• Identification of mental disorders • Differential diagnosis • Prognosis of treatment response or relapse
Unsupervised Learning	Data are unlabeled; the model independently identifies pat-terns	• Hierarchical clustering • k-means clustering • Principal component analysis (PCA) • Canonical correlation analysis	• Subtyping of mental disorders • Detection of clinical and neurobiological features
Semi-Supervised Learning	Combines labeled and unla-beled data	• Laplacian regularization • Semi-supervised clustering	• Multimodal analysis • Disease classification and diagnosis • Prediction from incomplete data
Deep Learning	A subset of ML using multi-layered artificial neural net-works; may be supervised or unsupervised	• Convolutional neural networks (CNNs) • Deep autoencoders • Graph convolutional networks • Recurrent neural networks (RNNs) • Long short-term memory (LSTM) • Generative adversarial networks (GANs) • Multiple-instance learning • Multilayer feedforward neural networks	• Processing genetic, neuroimaging, and other high-dimensional data to identify novel interactions

 

A related approach is semi-supervised learning, which utilizes labeled and unlabeled data for model training [20, 26]. This method has certain risks, as inaccuracies in the labeled data may be replicated by the model [27]. Despite the predominance of supervised learning in medicine, recent studies have demonstrated the high efficacy of semi-supervised learning in diagnosing BD relapse based on audio recordings [28].

In unsupervised learning, the model identifies patterns without labeled data. This method is effective for detecting similarities among patients with different mental disorders [19, 20, 26]. Unsupervised clustering may be used for identifying disease subtypes and analyzing clinical and biological heterogeneity in mental disorders [29, 30]. However, it poses challenges in understanding classification mechanisms and validating the accuracy and reliability of findings [31].

DL is a type of ML that employs artificial neural networks with multilayer architecture consisting of interconnected nodes across multiple layers [10, 17]. Several techniques have been developed to apply DL to datasets with <10,000 people [32]. The increasing availability of genetic, neuroimaging, and other datasets [33] further supports its use in psychiatry [32]. Some studies indicated that DL may offer superior effectiveness and accuracy in identifying novel associations and neurobiological mechanisms underlying mental disorders and their classification compared with traditional ML methods [34–36]. However, the interpretability and transparency of DL outputs remain limited, as the logic behind model decisions is often inaccessible [37]. Thus, some experts oppose the clinical use of DL in medicine [38].

The development of the ML model involves a series of steps to convert raw data into clinically meaningful outputs, such as a mental illness diagnosis [39–41]. The first stage includes data collection and preprocessing, comprising exploratory analysis, correlation assessment, class imbalance evaluation, imputation of missing values, noise and artifact reduction [42], selecting the most predictive variables [41], and transforming data using techniques appropriate to their type, such as natural language processing (NLP) for text data [36]. The next phase involves model training using one or more ML algorithms. Regardless of the method, data are typically divided into three subsets:

• Training set for model development;
• Validation set for preliminary performance assessment;
• Test set for final evaluation.

If the model inadequately performs during testing, researchers may refine it by expanding the dataset, revising hypotheses or feature sets, or modifying algorithms [21]. Once a satisfactory model is obtained, it is externally validated on a large, independent, representative clinical sample [43].

NEUROIMAGING DATA

Electroencephalography (EEG) and magnetic resonance imaging (MRI) are the most commonly used neuroimaging modalities in studies employing ML techniques, [10]. These imaging results are used to develop models for identifying high-risk individuals, diagnosing and classifying mental disorders, and predicting treatment outcomes [44].

Most studies focused on diagnosing mental disorders using supervised classification methods, with support vector machines (SVMs) being the most widely applied [44]. Gao et al. [25] reviewed 63 studies that utilized MRI data and ML methods to identify major depressive disorder biomarkers. The diagnostic accuracy of models based on structural MRI ranged from 68% to 90%, whereas models using functional MRI (fMRI) data showed an accuracy of 56%–100%. The highest accuracy was reported by Zeng et al. [45], who assessed resting-state fMRI data in 24 patients with severe recurrent depressive disorder (RDD) and 29 healthy controls, achieving 100% sensitivity using an SVM-based classifier. However, most models were tested on small samples, potentially limiting the reliability of their accuracy [44]. EEG data are also widely used for diagnosing depression. In a systematic review of 36 studies, Liu et al. [46] examined ML applications using EEG datasets for depression diagnosis, with model accuracy of 76%–99.5%. With MRI studies, sample sizes were generally small. The largest study included 400 participants (200 diagnosed with RDD and 200 healthy controls), wherein SVM training yielded test and training accuracy rates of 84.16% and 91.07%, respectively [47].

Claude et al. [48] analyzed 47 studies, 25 of which used ML methods to classify patients with BD and healthy controls, with reported accuracy of 57%–100%. Among these, 14 studies used only structural MRI data: 9 analyzed gray matter features (accuracy: 50%–90.7%), whereas 4 included both gray and white matter (accuracy: 63%–73%). Functional MRI data were evaluated in 8 studies (accuracy: 52.8%–83.5%), and only 2 studies used diffusion tensor imaging (accuracy: 78.12% and 100%) [49, 50]. Jie et al. [51] proposed a method for biomarker identification using SVMs with a forward/backward feature selection applied to structural and resting-state fMRI data. Discriminating features from both modalities enabled differentiation between BD and major depressive disorder, with an accuracy of up to 92.1%. Sample sizes of the reviewed studies ranged from 23 to 3020, with only 2 multicenter studies including >1000 participants. The models developed in these large-sample studies achieved a maximum accuracy of 65.23%, which was below the overall average model accuracy of 66%, with an area under the curve (AUC) of 0.6653 [48, 52, 53]. Four of the reviewed studies focused on identifying individuals at high risk for BD [48]. The highest performance was reported by Lin et al. [54], whose model distinguished between asymptomatic individuals with high genetic risk for BD (n = 34) and high-risk individuals with subsyndromal symptoms (n = 38), including hypomania, depression, psychosis, and attention-deficit/hyperactivity disorder, achieving an accuracy of 83.21% based on MRI-derived biomarkers.

Moreover, several studies applied ML techniques for differential diagnosis of mental disorders. Claude et al. [48] identified 11 studies that distinguished BD from schizophrenia, with model accuracy of 50%-96.2%. Sixteen studies attempted to differentiate BD from recurrent depressive disorder (RDD), reporting accuracy of 49.5%–93.1% [48]. The highest accuracy (90%) was achieved by a model trained on fMRI-derived patterns, although this pilot study included only 10 patients each with BD and RDD and 10 healthy controls, which may explain the increased performance [55]. Vai et al. [56] identified biomarkers to distinguish unipolar from bipolar depression using multimodal neuroimaging techniques, achieving an accuracy of 73.65%.

An emerging area of interest is the use of neuroimaging data to predict treatment outcomes. A meta-analysis by Watts et al. [57] published in 2022 included 15 studies employing ML and EEG-based datasets to forecast antidepressant response in patients with RDD. The cumulative accuracy of all models was 83.93% (95% CI: 78.90%–89.29%). Subgroup analysis revealed the highest predictive accuracy for rhythmic transcranial magnetic stimulation (cumulative accuracy: 85.70% [95% CI: 77.45%–94.83%]) and antidepressant therapy (cumulative accuracy: 81.41% [95% CI: 77.45%–94.83%]). Moreover, MRI data have been used to predict treatment response [44]. Liu et al. [58] analyzed fMRI data from 35 patients with RDD (18 with treatment-resistant depression and 17 responders) and 17 healthy controls. The accuracy of models assessing treatment efficacy was 85.7%–91.2%, depending on the evaluated parameters. However, the small sample size and high clinical heterogeneity limited the reliability of these results. Jiang et al. [59] studied electroconvulsive therapy (ECT) response based on MRI-derived gray matter volumes in patients with RDD and achieved a 90% accuracy. In a review by Claude et al. [48], only two studies applied ML methods to predict treatment response in BD. Wade et al. [60] used SVM classifiers trained on MRI data to predict ECT outcomes in patients with depressive episodes (45 with RDD and 8 with BD) and achieved an 89% accuracy. Furthermore, Fleck et al. [61] developed a classification model based on combined fMRI and proton magnetic resonance spectroscopy data to evaluate lithium treatment response, which achieved 88% and 80% accuracy in training and test sets, respectively (n = 20).

Some studies have employed unsupervised learning methods. Drysdale et al. [62] identified four neurophysiological subtypes of depression using resting-state fMRI. These were associated with distinct symptom profiles and treatment responses. Wu et al. [63] applied k-means clustering to define two homogeneous BD subgroups based on neurocognitive test results, MRI parameters, and clinical severities.

Fewer studies have focused on predicting recurrences or rehospitalization in mental disorders [10]. One study used a mixed dataset of 380 patients, including clinical, genetic, laboratory, and MRI data, to develop an SVM-based model predicting rehospitalization within 2 years after a first depressive episode, yielding an AUC of 67.74 [64].

Winter et al. [65] assessed the use of ML in identifying biomarkers of major depressive disorder. They analyzed data from 856 patients with RDD and 945 healthy controls and trained and tested 2.4 million ML models using neuroimaging data, polygenic depression risk scores, and clinical variables, including childhood trauma, social support, and medication load. Model accuracy did not exceed 62%, notably lower than in smaller studies. This result was primarily attributed to the large size of the test dataset, which limited the potential for artificial inflation of accuracy metrics, and decreased heterogeneity. In contrast to most studies, they used a wide range of ML algorithms, optimized hyperparameters, carefully selected predictors, and employed structured interviews to confirm RDD diagnosis. Notably, the study used only conventional ML methods. The authors noted that the nonlinear interactions that enhance DL model performance require datasets of >10,000 individuals; thus, DL was not likely to improve their results [66].

TEXT, AUDIO, AND VIDEO DATA

Thought and speech psychopathology is a core aspect of diagnosing most mental disorders [67]. One of the primary methods for converting speech, particularly the text content, into a format suitable for ML is NLP, which transforms text into numerical data [36]. NLP-based models are used to analyze patient data, including social media posts, transcripts of clinical interviews, and electronic medical records [67].

Early applications of ML to text data date back to 2013–2017, when researchers explored the detection of depression by analyzing social media activity [68–70]. Currently, dozens of such models exist [36]. In one study, an ML model predicted depression in 114/683 patients up to 6 months prior to formal clinical diagnosis based on social media activity (AUC = 0.72) [71]. In a review of 14 studies, Squires et al. [36] examined DL models trained on text datasets. The most accurate model identified depression based on social media posts, with 99% accuracy using convolutional neural networks (CNNs) and long short-term memory architecture [72]. Another study analyzed 27,308 posts from 146 users and reported 89% accuracy using CNNs; however, the diagnostic criteria for depression were not disclosed [73].

Data from electronic medical records have also been widely used to develop ML models for affective disorder diagnosis. Aggregate electronic medical record data from 4687 patients with RDD were used to predict rehospitalization using SVMs (AUC = 0.784) [74]. Edgcomb et al. [75] retrospectively analyzed electronic medical records from 552 patients with BD and used decision trees to predict 30-day post-discharge rehospitalization with 88% accuracy.

Additionally, transcripts of clinical interviews serve as input for ML models. One study analyzed transcripts from 1864 clinical interviews to assess suicide risk using ML, yielding an AUC of 0.82, comparable to clinician assessments [76].

Furthermore, audio data analysis extends beyond verbal content to include tone, volume, pitch, and vocal expressiveness [36]. Voice features vary in mental disorders [77]; for instance, increased voice tremor is observed in depression [78]. Using deep learning methods to analyze acoustic features in 189 audio recordings, a model diagnosed depression according to the Patient Health Questionnaire-8 (PHQ-8)³ with ~70% accuracy [79]. Another ML model classified depressive and mixed states (F1⁴ = 0.83 and 0.86, respectively) and hypomanic and mixed states (F1⁴ = 0.86 and 0.75) in 56 patients with BD based on audio features [80].

In addition to text and audio data, several studies have incorporated video data. Visual processing tools such as OpenFace can track facial landmarks, head position, facial muscle movement, and gaze [81]. Individuals with depression tend to smile and raise their eyebrows less frequently, frown more often, show reduced lip movement, and blink and gesture less [82, 83]. Using DL on video interviews, one model predicted depression severity based on the Beck Depression Inventory (MAE⁵ = 7.47; RMSE⁶ = 9.55) [84]. Another model analyzed gait from video recordings of 200 students and achieved an 85.45% accuracy [83].

Some studies have utilized multimodal datasets combining audio, video, and text data. Birnbaum et al. [85] developed a model using facial motion and audio recordings to differentiate schizophrenia (n = 41) from BD (n = 21), achieving an AUC of 0.73. A model developed using combined text and audio data (deep networks: one-dimensional CNN) for diagnosing depression (n = 189) demonstrated greater accuracy compared with models based on unimodal datasets (text or audio data alone): 0.91 vs 0.78 and 0.82, respectively [86]. Dibeklioglu et al. [87] used DL to create a model (n = 57) incorporating facial, head, and vocal dynamics to diagnose depression. The model trained on a multimodal dataset demonstrated higher accuracy (79%) than models based on unimodal data.

DATA FROM ELECTRONIC DEVICES

Data collected from wearable smart devices (e.g., smartwatches and fitness trackers) and smartphones may be valuable for assessing patients with affective disorders [88]. These electronic devices can track physical activity, geolocation, and usage metrics such as number of SMS messages, call logs, Internet activity, online purchases, music streaming, image viewing, and calendar use. Wearable devices passively collect physiological data, including heart rate, heart rate variability, electrodermal activity, body temperature, blood pressure, actigraphy, and sleep phase metrics [89]. In a systematic review, Seppälä et al. [90] analyzed 33 studies that evaluated sensor-derived data from electronic devices worn by individuals with schizophrenia, BD and RDD and subclinical populations experiencing depressive and anxiety symptoms. Nine studies identified associations between symptom severity (depression and mania) and measures of physical activity, smartphone usage time, geolocation, and typing behavior among patients with BD. Only two studies involved patients with clinical depression and 18 subclinical participants with depressive or anxiety symptoms [90]. The authors found correlations between depressive symptom severity and indicators such as geolocation, physical activity, and smartphone usage time [91, 92]. Considering their breadth, variety, and accessibility, sensor data from electronic devices are well suited for ML applications. A 2022 systematic review reported that ML models using unimodal datasets to diagnose affective disorders achieved accuracy of 70%–91% (n = 9). In contrast, models trained on multimodal datasets demonstrated higher accuracy (76%–98%) (n = 6) [89].

Several studies have specifically used smartphone-derived data to diagnose affective disorders. Opoku Asare et al. [93] created datasets from 629 participants that included battery usage, time zone, app and Internet usage, screen-locking behavior, and demographic variables. The best-performing ML algorithm trained on these datasets achieved a 92.51% accuracy. However, the authors noted that depression was assessed using the PHQ-8³, a self-reported measure, without clinical validation; thus, model accuracy may differ in clinically diagnosed populations. In another study using the random forest method, a model based on call duration and text message count achieved 81.1% accuracy in diagnosing bipolar depression (n = 412) [94]. A model using app usage data (e.g., email, social media, video, audio, gaming, shopping, and education) predicted depression with 80% accuracy (n = 79) [95].

Geolocation data can also be informative. One study used a dataset tracking the geolocation changes among 28 patients to develop an ML model that identified depressive phase onset with an 86.5% accuracy [96].

Actigraphy data are among the most frequently used for developing models aimed at early detection of affective phases in RDD and BD. One study used actigraphy data collected over 1 week to develop a model differentiating melancholic depression (n = 8) from other depressive subtypes (n = 7), achieving an AUC of 0.84 [97]. Using adaptive boosting and 2-week actigraphy datasets, another model distinguished patients with depression (15 with RDD and 8 with BD) from healthy controls (n = 32), reaching 78% accuracy [98]. A model trained using random forest on 90-day actigraphy data from 25 BD patients and 25 controls achieved 88% accuracy, with 85% sensitivity and 91% specificity [99]. In the largest study involving data from patients with RDD (n = 24,229) and healthy controls (n = 4124), the classification model achieved an AUC of 0.68 (95% CI: 0.67–0.69). The model also distinguished between patients with typical (n = 18,722) and atypical (n = 958) depressive symptoms (e.g., hypersomnia and weight gain), with an AUC of 0.74 (95% CI: 0.71–0.77). Key predictors included difficulty waking, insomnia, snoring, daytime inactivity, and reduced activity at 8:00 a.m. [100]. Jakobsen et al. [101] used deep learning on actigraphy data from 23 patients with BD or RDD and 32 controls to develop a classification model with 84% accuracy [101].

An inverse correlation was found between heart rate variability and both depression severity and rumination intensity [102], possibly due to increased parasympathetic nervous system activation [103]. This physiological basis has been used in ML model development. Byun et al. [104] employed SVMs with heart rate variability data from 33 RDD patients and 33 controls to create a depression detection model with 70% accuracy. Another study achieved 86.4% accuracy using the Bayes classifier based on heart rate variability metrics for diagnosing RDD [105].

Some studies have used multimodal data from electronic devices [89]. The largest of these was conducted by Nickels et al. [106], which involved 415 participants (approximately 80% with RDD). Using a smartphone app, the authors assessed 34 variables, 11 of which significantly correlated with PHQ-9⁷ total scores. These variables included:

• Audio diary characteristics: duration, number of pauses, and mood-related word usage
• Sleep duration
• Ambient noise levels
• Number of calls and response latency
• Geolocation data: novelty, variety of locations, and time spent at home
• Battery charge level
• Battery charging patterns
• Emoji and mood-related word use in text messages
• Screen time
• Volume settings
• Frequency of social app usage.

A logistic regression model based on these features showed a mean AUC of 0.656.

In a smaller study of 20 participants with varying PHQ-9⁷ depression scores, ML models using SVM and random forest methods classified depression categories with 96% accuracy. Smartphone and smartwatch apps were used to assess five symptom clusters: physical activity, social activity, mood, sleep, and eating behavior [107]. Cho et al. [108] conducted a prospective observational cohort study of 55 patients with major depressive disorder and BD types I and II. The participants used a smartphone app for self-reporting daily mood scores, while a sensor tracked light exposure. Activity trackers recorded digital logs of activity, sleep, and heart rate. The authors developed a mood prediction algorithm using the random forest method, which achieved 65% accuracy in forecasting mood changes over the following 3 days. The model accuracy in detecting clinical states of remission, depressive, manic, and hypomanic episodes was 85.3%, 87%, 94%, and 91.2%, respectively.

Laboratory and Molecular Genetic Data

Omics technologies are one of the fastest-growing areas in precision medicine, particularly in psychiatry. An increasing number of molecular biomarkers have been associated with mental disorder onset and therapeutic response. The advent of next-generation sequencing has further accelerated the development of bioinformatics and ML methods for managing these large-scale datasets. ML techniques can integrate pharmacogenomic, epigenetic, metabolomic, transcriptomic, and proteomic data to map specific neurobiological substrates to symptom clusters, enabling refined diagnostic classifications, personalized treatment selection, and outcome prediction [40].

Single-nucleotide polymorphisms (SNPs) are a most frequently used feature for training ML models. In one study, SVMs predicted remission with duloxetine therapy at 52% accuracy, with 58% sensitivity and 46% specificity (n = 186) [109]. Another model using random forest algorithms predicted 8-week remission in patients treated with citalopram or escitalopram (n = 398), achieving a 69% accuracy (AUC > 0.7) [110]. Eugene et al. [111] developed models using decision trees and random forests to predict lithium response in BD. Predictors included RBPMS2 and LILRA5 expression in males and ABRACL, FHL3, and NBPF14 expression in females. The models achieved sensitivities of 96% and 92%, respectively.

Qi et al. [112] used microRNA profiles from 168 patients with RDD as predictors. Their ML model (supervised classification) differentiated depressed individuals from healthy controls (AUC = 0.97) and mild from severe depression according to the Montgomery–Asberg Depression Rating Scale (AUC = 0.63) and predicted treatment response (AUC = 0.57).

Some studies employed multimodal datasets combining genetic, demographic, and clinical features for diagnostic model development. Lin et al. [113] studied 455 RDD patients to build DL models predicting selective serotonin reuptake inhibitor response. Multilayer feedforward neural networks incorporated the following:

• 10 SNPs
• Demographic variables
• Hamilton Depression Rating Scale scores
• Number of affective episodes
• History of suicide attempts.

A model with two hidden layers achieved an AUC of 0.82 (75% sensitivity and 69% specificity). Another model integrating neuroimaging, genetic, DNA methylation, and demographic data (n = 121) reached 84% accuracy [114].

Five studies investigated ML-based diagnosis of BD using molecular genetic data [115]. A study applied various classification methods to genome-wide association study datasets from 2191 patients with BD and 1434 volunteers. The best-performing model was the Bayes classifier (AUC = 0.56) [116]; all other methods underperformed compared to polygenic risk score approaches [117]. A higher accuracy (74%) was reported using a random forest classifier, although the dataset was smaller (604 patients with BD and 1767 healthy volunteers) [118]. Laksshman et al. [119] used CNNs and BD genotypic data to develop a DL model (DeepBipolar), which achieved an AUC of 0.65.

Laboratory indicators are less commonly used as predictors in ML models. Wollenhaupt-Aguiar et al. [120] developed ML models to classify patients with bipolar depression (n = 54) and unipolar depression (n = 54), achieving an AUC of 0.69 (62% sensitivity and 66% specificity). Predictors included three biomarkers: interleukin-4, thiobarbituric acid-reactive substances, and interleukin-10. A separate model differentiated patients with bipolar depression from healthy controls using five biomarkers: interleukin-6, interleukin-4, thiobarbituric acid-reactive substances, carbonyl compounds, and interleukin-17A, with an AUC of 0.70 (62% sensitivity and 70% specificity). Another model comparing patients with unipolar depression and healthy controls used seven variables: interleukin-6, carbonyl content, brain-derived neurotrophic factor, interleukin-10, interleukin-17A, interleukin-4, and tumor necrosis factor-alpha. This model achieved an AUC of 0.74 (68% sensitivity and 70% specificity).

ADVANTAGES AND LIMITATIONS OF MACHINE LEARNING IN AFFECTIVE DISORDERS

This review summarizes key studies focused on ML model development for patients with RDD and BD, using predictors derived from neuroimaging, text, audio, and video data; wearable devices; omics biomarkers; and laboratory parameters. These models can be applied across various stages of psychiatric care, including:

• Identifying at-risk populations to enhance early detection;
• Diagnostic support including severity, subtype differentiation, and disease trajectory;
• Personalizing treatment selection for optimal response;
• Predicting therapeutic outcomes;
• Forecasting relapse risk.

Fig. 1 outlines the data types used, ML model development stages, and clinical applications in affective disorders.

 

Fig. 1. Stages of machine learning development and implementation in clinical practice for patients with affective disorders. DNA, deoxyribonucleic acid; DTI, diffusion tensor imaging; EEG, electroencephalography; GPS, Global Positioning System; ML, machine learning; MRI, magnetic resonance imaging; RNA, ribonucleic acid; SNPs, single-nucleotide polymorphisms.

 

A primary advantage of AI in psychiatry is its ability to incorporate heterogeneous data types. Integrating multiple modalities may improve model accuracy and provide deeper insights into inter-variable relationships [121]. Recently, substantial efforts have been dedicated to developing ML methods for constructing multimodal datasets [10]. Only a few studies applied such models [55, 86, 87], with some reporting higher accuracy using multimodal datasets [86, 87]. However, Winter et al. [65], after analyzing 2.4 million ML models, argued that multimodal data integration does not necessarily enhance diagnostic or predictive accuracy. They attributed this to the models having very few high-quality predictors to benefit from integration or a strong correlation among the included variables. Hence, there remains a need for interpretable ML methods that can effectively combine data types and be validated in large-scale studies.

Another advantage of AI is its independence from direct clinician involvement, allowing for diagnostic and treatment guidance even before physician consultation. By the time of the visit, clinicians can access crucial information such as probable diagnoses and optimal treatment options, potentially improving diagnostic accuracy, reducing consultation time, and mitigating workforce shortages [122]. Furthermore, AI systems are not subject to human factors such as faulty operation, stress, and fatigue [18]. Nonetheless, AI cannot replace qualified clinicians owing to several limitations.

A major limitation in the reviewed studies is small sample size, which increases the risk of model overfitting and may artificially inflate accuracy metrics. Another issue is the limited representativeness of study cohorts, hindering generalizability. Furthermore, the lack of standardization in data collection and processing can affect results and reduce reproducibility of studies [123, 124]. Most studies relied on self-report questionnaires to establish depression rather than structured diagnostic interviews, which may limit the applicability of models to clinically diagnosed populations. Inclusion of irrelevant or noisy features may also distort model performance. Model accuracy can improve if training and testing datasets contain relevant correlations, even if such associations are not immediately evident (e.g., comorbid conditions) [43]. Another key drawback is the lack of external validation, which is critical before ML models can be translated into clinical practice [125]. A robust path toward AI technology implementation should include internal validation, independent dataset validation, and testing in randomized controlled trials [126].

These concerns highlight the urgent need for standardized research protocols and quality control prior to clinical deployment of AI tools [127]. Additionally, there is a critical demand for large, multimodal datasets, including neuroimaging, genetic, clinical, audio, and video data. The number of such databases is steadily increasing. The ENIGMA⁸ database, developed by over 500 researchers from 35 countries, includes neuroimaging (>30,000 MRI scans), genetic, and clinical data of patients with brain disorders [33]. These advances raise ethical challenges. Researchers should ensure data confidentiality, obtain informed consent, and communicate clearly how patient data will be used. Establishing uniform ethical guidelines for AI in medicine [128], especially in psychiatry, remains urgent [129, 130].

Moreover, ML application in psychiatry is challenged by several conceptual barriers, including the lack of clear diagnostic criteria for mental disorders, subjective symptom assessment, low specificity, and high clinical heterogeneity at the individual level [10]. These factors hinder the development of unified algorithms that could be a foundation for ML models to diagnose mental disorders, unlike in other areas of medicine [131]. Shifting the focus from diagnosis to prognosis and disease outcomes, along with alternative data processing approaches and ML models capable of uncovering novel associations, may improve the accuracy of results [65]. Moreover, normative modeling approaches that emphasize individual-level characteristics [132], dynamic systems theory [133], and network theory of mental disorders [134] may represent emerging directions for integrating AI technologies into managing affective disorders [30]. However, the effectiveness of these approaches remains unconfirmed based on current data.

This review had several limitations, the most significant being the lack of a systematic analysis with statistical data processing, primarily due to the absence of a homogeneous dataset on the use of AI technologies in patients with affective disorders. Inclusion criteria were intentionally broad to capture a wide range of studies; however, this may have resulted in omitting relevant publications. The review solely focused on studies involving patients with RDD and BD and did not address the application of AI in psychiatric research planning, hypothesis generation [135], mobile health applications [136], neurostimulation with feedback mechanisms [137], or Chat Generative Pre-Trained Transformer technologies [138].

CONCLUSION

This review summarized current approaches in ML and DL, stages of model development, and key studies applying ML to the diagnosis of patients with RDD and BD, using various data as predictors. The most commonly used data types in ML models for affective disorders include neuroimaging (MRI and EEG), text, audio, and video data; data from electronic devices; and genetic, clinical, and laboratory variables and their combinations.

ML methods have demonstrated promising potential in the early detection of affective episodes and in predicting long-term therapeutic responses to mood stabilizers [139]. However, the implementation of AI technologies in clinical psychiatry remains limited owing to multiple technical and conceptual challenges. Technically, the limited quality of models is attributed to small sample sizes, low representativeness and standardization, inclusion of noise (extraneous factors) and intercorrelated variables, and lacking validation in independent patient cohorts. Additionally, the lack of objective assessment in psychiatry may hinder the development of accurate models using ML methods. Improving AI algorithm performance requires the creation of large, high-quality datasets, development of standardized research quality metrics, and ethical guidelines for AI use. There is also a need for ML models that leverage multimodal data and can identify novel inter-variable relationships. This demands close collaboration among computer scientists, researchers, and clinical practitioners. Additionally, it is critical to foster public and patient engagement in the evolution of AI technologies and enhance awareness of their potential benefits and limitations.

ADDITIONAL INFORMATION

Funding source. This article was carried out within the framework of the topic of the state assignment by State grant of V. Serbsky National Medical Research Centre for Psychiatry and Narcology of Ministry of Health of the Russian Federation, (USIS No 124020800062-5).

Disclosure of interests. The authors declare that they have no relationships, activities or interests (personal, professional or financial) with third parties (commercial, non-commercial, private) whose interests may be affected by the content of the article, as well as no other relationships, activities or interests over the past three years that must be reported.

Author’s contribution. E.S. Mоsolova: concept of the work, collection and analysis of literary data, writing the manuscript; A.E. Alfimov: scientific editing of the manuscript; E.G. Kostyukova editing of the manuscript; S.N. Mosolov: concept of the work, editing of the manuscript. Thereby, all authors provided approval of the version to be published and agree to be accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

¹ Institute of Health Metrics and Evaluation [Internet]. Global Health Data Exchange (GHDx); 2021. Available at: https://vizhub.healthdata.org/gbd-results Accessed on February 14, 2024.

² Bestsennyy O., Gilbert G., Harris A., Rost J. Telehealth: a quarter-trillion-dollar post-COVID-19 reality? [approx 15 pages]. In: McKinsey & Company [Internet]. 2021–2023. Available at: https://www.mckinsey.com/industries/healthcare/our-insights/telehealth-a-quarter-trillion-dollar-post-covid-19-reality Accessed on: January 15, 2024.

³ PHQ-8 (Patient Health Questionnaire) is a version of the PHQ-9 depression screening tool that omits the final item on thoughts of death or self-harm. It is based on the diagnostic criteria from the DSM-5.

⁴ F1-score is a machine learning metric that provides a balanced measure of model performance by combining both precision and recall.

⁵ Mean Absolute Error is a metric that quantifies the average magnitude of errors between actual and predicted values.

⁶ Root Mean Squared Error is the square root of the mean of the squared differences between predicted and actual values. It is one of the primary metrics used to assess the performance of regression models.

⁷ PHQ-9 (Patient Health Questionnaire) is a questionnaire consisting of 9 items used to assess depression in accordance with the diagnostic criteria of the DSM-5.

⁸ ENIGMA [Internet]. The ENIGMA Consortium; 2009–2024. Available at: http://enigma.ini.usc.edu Accessed on February 14, 2024.

About the authors

Ekaterina S. Mosolova

V. Serbsky National Medical Research Centre for Psychiatry and Narcology

Email: kata_mosolova@mail.ru
ORCID iD: 0000-0003-2324-2814
SPIN-code: 6077-3386
Russian Federation, Moscow

Alexander E. Alfimov

Sechenov First Moscow State Medical University

Email: alex.alfimov@gmail.com
ORCID iD: 0000-0002-9064-7881
SPIN-code: 4354-7081

MD, Cand. Sci. (Medicine)

Russian Federation, Moscow

Elena G. Kostyukova

V. Serbsky National Medical Research Centre for Psychiatry and Narcology

Email: ekostukova@gmail.com
ORCID iD: 0000-0002-9830-1412
SPIN-code: 6510-3969

MD, Cand. Sci. (Medicine)

Russian Federation, Moscow

Sergey N. Mosolov

V. Serbsky National Medical Research Centre for Psychiatry and Narcology; Russian Medical Academy of Continuous Professional Education

Author for correspondence.
Email: profmosolov@mail.ru
ORCID iD: 0000-0002-5749-3964
SPIN-code: 3009-9162

MD, Dr. Sci. (Medicine), Professor

Russian Federation, Moscow; Moscow

References

Supplementary files