Mood disorders are a highly prevalent and recurrent group of mental disorders associated with a substantial risk of suicide [1], primarily encompassing unipolar depression (UD) and bipolar disorder (BD) [2]. Approximately 1 in 4 individuals is estimated to experience an affective disorder at least once in their lifetime, often leading to substantial and lasting disability for those affected [3]. UD is primarily characterized by substantial and persistent low mood. In contrast, BD involves manic or hypomanic episodes (elevated mood, racing thoughts, and increased activity) and depressive episodes (low mood, slowed thinking, and reduced activity). Both disorders are marked by enduring mood changes impacting emotional, cognitive, and behavioral domains [4]. However, the course of BD often begins with depressive episodes, leading to a substantial risk of misdiagnosis, as approximately 40% to 69% of patients with BD are initially diagnosed with UD [5,6]. This misdiagnosis can have serious consequences, including inappropriate medication prescriptions, triggering manic episodes, prolonged illness duration, increased risk of recurrence, heightened suicide risk, and an overall poorer response to treatment [5,7-10].

Currently, psychiatrists typically diagnose based on established criteria (such as the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition and the International Classification of Diseases, Eleventh Revision), relying on one-time self-reports from patients and their families. This approach heavily depends on the clinician’s experience and often lacks continuity and objectivity. In contrast, digital phenotyping offers a promising solution, with this concept being a groundbreaking advancement in the field, first introduced in 2015 [11]. Digital phenotyping includes long-term active data (eg, participants completing daily self-assessment questionnaires via smartphone apps) [12], providing clinicians with a more comprehensive and continuous flow of information. It also offers objective data based on digital devices, such as physiological measurements (eg, skin temperature, heart rate, blood volume pulse, and electrodermal activity) and behavioral indicators (eg, acoustic features, gestures, and facial expressions) [13].

Compared to traditional diagnostic methods, digital phenotyping has great potential to improve diagnostic accuracy and timeliness. However, it generates vast amounts of data that require more robust processing and analysis. To address this, leveraging artificial intelligence, such as machine learning, in mental health is essential. Currently, machine learning has gradually emerged as a powerful tool for exploring high-dimensional and real-time data associated with digital phenotyping. It provides an opportunity to “make sense” of these digital phenotypes and the realities they attempt to represent in the context of mental health [14-16]. Some believe this technology has the potential to offer deeper insights into the neurobiological mechanisms underlying psychiatric disorders [17]. It may also facilitate the development of new transdiagnostic models for understanding symptoms [12]. This aligns with the “research domain criteria” perspective proposed in recent years [18]. In addition to classification tasks, machine learning algorithms may also have the capability to predict episodes or even suicidal risk, enabling clinicians to make more accurate and timely clinical decisions. Consequently, digital phenotyping, supported by machine learning, has carved out an important role in psychiatry [19], helping clinicians access individualized behavioral, emotional, and other data from patients with mental disorders. This approach not only enhances psychiatrists’ understanding of symptoms and the disorders themselves [20,21], but may also compensate for the current lack of reliable biomarkers.

In this study, we conducted a systematic review of original articles from both journals and conference proceedings, exploring the use of portable or wearable digital tools for distinguishing UD and BD, as well as classifying UD, BD, and healthy control (HC) individuals. We examined the studies by considering factors such as demographic characteristics and the diagnostic criteria or psychiatric assessments used. In addition, we analyzed the technological tools and data types used, the duration of data collection, data preprocessing methods, selected variables or features, computational techniques applied, validation approaches, and the results achieved.

This review is essential because many previous studies in this field have either focused on specific tools or were limited by small sample sizes, which restricts their applicability to broader populations. Furthermore, the lack of consistency in methodologies and the absence of comprehensive validation across different groups have hindered the generalizability of the findings. This review aimed to address these limitations and provide valuable insights into the potential of digital phenotyping for more accurate and reliable differentiation between these groups.

This systematic review was conducted in agreement with the PRISMA (Preferred Reporting Items for Systematic reviews and Meta-Analyses) guidelines (Multimedia Appendix 1) [22] and is listed in the PROSPERO register (CRD42024624202). While the review was generally conducted according to the registered protocol, some minor deviations occurred during the study. These adjustments primarily involved refining and clarifying the study objectives and title to better reflect the focus of the review. In addition, we expanded the scope to include studies that classified UD, BD, and HC, which was not initially specified in the protocol. Finally, we excluded studies based on nonclinically diagnosed data, such as those relying on social media samples, to ensure that only studies adhering to professional diagnostic standards were included.

We conducted a comprehensive search across 6 major databases, including Scopus, IEEE Xplore, PubMed, Embase, Web of Science, and PsycINFO, for articles published up to March 20, 2025. The search terms used in each database are provided in Multimedia Appendix 2. We included original studies, both journal articles and conference papers, published in English, with no restrictions on publication date.

Eligible studies involved participants diagnosed with UD, BD, or HC and used portable or wearable digital devices such as smartphone apps, wearable sensors, or audio or visual recordings. The studies were required to either compare digital phenotyping results with diagnostic outcomes from professional medical evaluations, compare UD with BD, or perform a classification task involving UD, BD, and HC. The primary outcome of interest was classification performance metrics such as sensitivity, specificity, area under the curve (AUC), accuracy, recall, and precision, but studies reporting t tests, ANOVA, nonparametric tests, or correlation analyses were also considered.

Following the PICOS (population, intervention, comparison, outcome, and study design) criteria, we excluded studies that were reviews, meta-analyses, or written in languages other than English. In addition, studies were excluded if they used technologies unsuitable for daily monitoring, were based on electronic health records or clinical data, or had diagnoses not made through professional medical evaluations (eg, studies based on social media data).

The included papers were grouped based on the type of digital tool used (smartphone apps, wearable devices, audiovisual recordings, or multimodal technologies) and the comparison type (UD vs BD or UD vs BD vs HC). A narrative synthesis approach was used to summarize and compare study characteristics, data modalities, analytical methods, and classification performance. Quantitative pooling was not performed due to the heterogeneity in study designs and feature selection.

The quality of the selected studies was independently evaluated by 2 reviewers (RZ and XW) using the Quality Assessment of Diagnostic Accuracy Studies-2 (QUADAS-2) [23] tool. Discrepancies were resolved through discussion or with the involvement of a third author (YF). QUADAS-2 comprises 4 key domains: patient selection, index test, reference standard, and flow and timing. For each domain, the risk of bias and concerns regarding applicability were assessed and categorized as low, high, or unclear risk.

As shown in Figure 1, a thorough search was conducted across 6 major databases, yielding a total of 2555 articles. Specifically, PsycINFO identified 90 articles, PubMed provided 114 articles, IEEE Xplore returned 231 articles, Embase found 356 articles, Web of Science collected 782 articles, and Scopus gathered 982 articles. After removing 1013 duplicate records, the titles and abstracts of the remaining 1542 articles were screened for relevance to the topic, narrowing the selection to 89 articles for full-text eligibility assessment. During the eligibility screening, 68 full-text studies were excluded after a detailed evaluation. Among these, 35 studies applied digital phenotyping but did not distinguish between UD and BD or classify UD, BD, and HC. In total, 11 studies did not use digital phenotyping at all. Another 11 studies used digital phenotyping to track emotional states or make predictions. The remaining 11 studies were either protocols, introductions to machine learning methods, or reviews. These exclusions ensured that the final selection aligned precisely with the study’s specific objectives. The final sample comprised 21 articles, with the earliest study dating back to 2016, emphasizing the novelty of this research topic. Of the 21 studies, 11 (52%) studies directly distinguished between UD and BD (Table 1 and Multimedia Appendix 3 [24-44]), while 10 (48%) studies classified UD, BD, and HC (Table 2 and Multimedia Appendix 3). It is important to note that studies based on the RADMIS trials or those using the CHI-MEI mood disorder database may have partly overlapping populations. However, they were not excluded due to their very different methods and results. In addition, although studies using the CHI-MEI database did not explicitly report the diagnostic criteria or psychiatric assessments, data collection was carried out in collaboration with clinicians at Chi Mei Medical Centre. Consequently, these studies were not excluded.

The results of the methodological quality assessment using the QUADAS-2 tool were presented in Multimedia Appendix 4. The risk of bias in the included studies is considerable and cannot be overlooked. A detailed table summarizing the QUADAS-2 scores for each study is provided in Multimedia Appendix 5 [24-44].

We conducted a systematic review of original articles from both journals and conference proceedings, investigating the use of portable or wearable digital tools for distinguishing between UD and BD, as well as for classifying UD, BD, and HC. A total of 21 articles were included, categorized into 4 main groups based on the type of digital tool assessed: (1) smartphone apps for collecting active data such as mood self-assessments or passive data, location, naturalistic phone calls, device rotation, and more; (2) wearable sensors, including the Actiwatch and the research-grade wearable device Empatica E4; (3) audiovisual recordings for analyzing speech characteristics, facial expressions, and upper body movements; and (4) multimodal technology, which combine text, audio, facial expressions, heart rate, and eye movement data. Overall, digital phenotyping data offer substantial opportunities for advancing the differential diagnosis of mood disorders. Despite certain methodological limitations, our findings highlight the potential of these digital technologies to provide more precise and objective support in diagnosing mood disorders. Certain features, such as activity levels captured through smartphone apps or wearable devices, emerged as potential markers for directly distinguishing between UD and BD. Individuals with BD typically exhibited lower activity levels compared to those with UD. Moreover, patients with BD tended to show higher activity levels in the morning and lower activity in the evening, whereas patients with UD displayed the opposite pattern. In addition, approaches leveraging speech modalities or integrating multiple modalities achieved better classification performance across the UD, BD, and HC groups, although the specific contributing features remain unclear. This uncertainty could be attributed to the complex and diverse nature of voice features, including strictly acoustic features (eg, speech rate, pause duration, and response time) [33], prosodic features (eg, pitch, energy, and formants) [43,46], and spectral features (eg, gamma-tone frequency Cepstral coefficients and Mel-frequency cepstral coefficients) [34,47].

Diagnostic confusion between BD and UD often occurs when patients are in a remitted or depressive state, as patients and their families may fail to recall previous manic or hypomanic episodes, making it challenging for clinicians to make an accurate diagnosis [24]. Of the 21 included studies, 6 (29%) included studies that used smartphone apps to directly distinguish between UD and BD. Of these, 5 (83%) studies using the Monsenso system considered the clinical staging of both groups and used different types of digital phenotyping data to differentiate between UD and BD. Among them, 3 (60%) studies used machine learning algorithms, with the best AUC of 0.79 achieved in both euthymic and depressive states (based on location data) [24]. However, results from 2 (40%) other studies [25,28] showed lower AUC values (0.42-0.58), which may indicate that variations in smartphone-based digital phenotyping are highly individualized [28]. In total, 4 (67%) [25-27,29] out of 6 studies used participants’ active data to distinguish between UD and BD, including daily self-assessments of irritability, mood, activity, sleep, stress, anxiety, and naturalistic phone calls voice. Half of these studies (2/4, 50%) used mixed-effects models, revealing differences in the level of activity [29] and presence of irritability [27] between UD and BD in a depressive state.

Interestingly, among the 3 (14%) out of 21 studies that used wearable devices to collect digital phenotyping data, 1 (33%) study also revealed differences in activity patterns between UD and BD. In contrast to the study mentioned earlier that used daily self-report activity scales, this study collected participants’ activity levels using activity trackers. This study provided a more detailed insight into the activity patterns of UD and BD: BD showed a high activity pattern in the morning and a low activity pattern in the evening, while UD exhibited the opposite [30]. Another study based on wearable watches also considered activity levels as a digital phenotype, but included HC as well. Using machine learning for a 4-class classification of UD, BD type 1, BD type 2, and HC, the study achieved an accuracy of 0.991 [36]. These 3 studies all demonstrate the potential of activity levels in the differential diagnosis of mood disorders.

Among the studies using audiovisual recordings and multimodal technologies to classify UD, BD, and HC, nearly all involved one-time data collection from participants in a standardized environment. Overall, approaches based on speech modalities or the integration of multiple modalities demonstrated better classification performance across these 3 groups, with accuracy ranging from 75.56% to 95.6% [37,40,41,43,44]. However, the specific features contributing to the classification performance remained unclear.

Accordingly, it can be concluded that smartphone apps and wearable devices, when combined with machine learning or other advanced analytical methods, demonstrate moderate diagnostic potential for differentiating UD and BD. However, their effectiveness is often limited by substantial individual variability in digital phenotyping data. Similarly, studies using audiovisual recordings and multimodal technology with machine learning have shown mixed outcomes, while they exhibit promising results in multiclass classification tasks, they have had limited success in directly distinguishing between UD and BD. These findings suggest that, although digital phenotyping holds potential for supporting the differential diagnosis of mood disorders, it should complement, rather than replace, comprehensive clinical evaluations conducted by trained professionals.

In particular, digital phenotyping data collected via smartphones or wearable devices or sensors may be more suitable for long-term dynamic monitoring and even intervention in the real world. This advantage lies in its ability to continuously gather active or passive data, offering a more comprehensive understanding of an individual’s mental and physical state. By capturing data in real time, these approaches help mitigate recall biases commonly associated with other symptom-reporting methods [48]. Moreover, they address the issue of “back-filling,” a frequent problem with paper-based diaries [49], thereby improving the reliability and accuracy of symptom tracking. However, factors such as patients’ proficiency in using electronic devices, the need for continuous and uninterrupted use, and potential limitations in monitoring continuity due to lack of feedback must be considered [50], as these could impact its medical value. In contrast, multimodal data collected through fixed paradigms in controlled settings offer advantages, such as capturing explicit behaviors (eg, speech, facial expressions, gestures) and implicit physiological signals (eg, eye movements, heart rate). Standardized protocols ensure reliable and reproducible data, making this structured approach particularly effective for mood disorder detection and classification. For instance, in the studies by Valstar et al [51] and Ringeval et al [52], audiovisual signals have been proven to play a substantial role in the detection and classification of UD and BD.

Our systematic review primarily focused on criterion validity and content validity. Criterion validity, defined as the extent to which the results of a specific test align with those of a reference standard [53], was demonstrated in the included studies by comparing classifications or distinctions based on digital phenotyping with diagnostic outcomes from professional medical evaluations. Content validity, on the other hand, refers to the degree to which an assessment tool is relevant to and representative of the targeted construct it aims to measure [54]. The included studies primarily focused on capturing one or a few dimensions of emotions (eg, daily mood, social interactions, or activity levels), falling short of comprehensively reflecting all core features of mood disorders. Furthermore, the complexity and vastness of digital phenotyping data led most studies to use machine learning algorithms for classification tasks, which posed substantial challenges for the evaluation of discriminant validity and structural validity.

The article that first introduced the concept of digital phenotyping stated that data collected through social media, forums, online communities, wearable technologies, and mobile devices offers substantial value that goes beyond traditional methods such as laboratory tests and clinical imaging [11]. This is particularly substantial given the unclear pathogenesis, the unknown etiologies, and the lack of objective biomarkers of mood disorders. As a result, digital phenotyping stands out as a promising tool in supporting mood disorder diagnosis. We should also recognize that as medical technology advances, the types of digital phenotyping data are likely to expand. Beyond the digital devices discussed in our review, smart speakers may also play a role in future applications [55]. Besides, traditional laboratory tests and imaging methods are becoming more portable, including smartwatch-based 9-lead electrocardiograms [56], wearable electroencephalography [57], and even devices for long-term monitoring of metabolic indicators, such as continuous glucose monitors [58]. This evolution suggests that, in the future, more complex and extensive multimodal data may be used for the auxiliary diagnosis of mood disorders. However, the growing diversity of data types increases the demand for more sophisticated data analysis and processing techniques. To meet these challenges, integrating multimodal data from various sources could provide a more comprehensive understanding of mood disorders and expand the potential applications of digital phenotyping in mental health.

Despite the promise of digital phenotyping, we observed that this field is still relatively new, with a limited number of studies conducted thus far. Most of the research in this area involves small sample sizes and lacks consistency in evaluation standards and procedures. This makes it difficult to draw generalizable conclusions and underscores the need for further research to refine methodologies, improve data quality, and establish robust analytical frameworks. Therefore, it is crucial to develop standardized methodologies, guidelines, and protocols in the field of digital phenotyping to ensure consistency and reliability in data collection, analysis, and interpretation. Establishing these standards will enable better comparisons across studies, improve reproducibility, and ultimately enhance the clinical applicability of digital phenotyping for the diagnosis and monitoring of mood disorders.

Undoubtedly, digital phenotyping plays a substantial role in the prevention, diagnosis, treatment, and management of mood disorders and even mental disorders in general. However, several key requirements must be met to transform the way mental health care is delivered [59]. First, given the unique nature of mental disorders, patient privacy and data security require greater attention than for the general population. Patients must be fully informed and consent to how their data are collected, managed, and used, as this is crucial to maintaining trust between patients and clinicians [16]. Moreover, the “black box” nature of machine learning models continues to raise long-term concerns among health care professionals [60], as these models obscure their decision-making processes [12]. Such opacity limits the clinical actionability of machine learning models, particularly for classification tasks, as clinicians often require clear explanations to trust and act upon this prediction. This underscores the importance of explainable artificial intelligence [61], which aims to make the reasoning behind machine learning conclusions transparent, thus enhancing trust among health care professionals and ensuring that patients maintain greater control over their health data. However, the integration of digital tools into mental health care also faces broader social challenges. For example, these technological advances may exacerbate existing inequalities in access to psychiatric care, particularly disadvantaging underprivileged populations who lack access to the necessary technology or infrastructure [62]. Furthermore, the stigma surrounding mental illness [63], which already acts as a substantial barrier to seeking help, may be compounded by the perceived invasiveness of digital health tools, discouraging their adoption by those most in need.

To the best of our knowledge, this is the first systematic review to comprehensively synthesize existing evidence on distinguishing BD from UD based on digital phenotyping. By systematically evaluating studies using smartphone apps, wearable devices, audiovisual recordings, and multimodal technologies, this review highlighted the emerging potential of digital phenotyping as a supportive tool for the differential diagnosis of mood disorders. However, our systematic review has several limitations. First, many of the included studies had small sample sizes, which increases the risk of overfitting and limits the generalizability of the findings. Future research should involve larger and more diverse samples to improve model robustness and applicability. Second, none of the studies validated their findings using external datasets, raising concerns about the reliability and reproducibility of the results. Incorporating independent datasets for external validation is essential to ensure consistent performance across different populations. Third, there was substantial heterogeneity in study designs, including variations in data sources, analytical methods, and outcome measures. This heterogeneity complicates direct comparisons and weakens the strength of synthesized conclusions. To address this, future studies should adopt standardized methodologies and outcome definitions to enhance comparability.

In conclusion, our review shows that digital phenotyping for distinguishing UD and BD is progressing rapidly. However, challenges such as privacy, data security, and equitable access must be addressed for digital health care to effectively transform mental health care. Only by overcoming these challenges can digital innovations fulfill their potential to improve mental health care inclusivity.