Wearable-Derived Behavioral and Physiological Biomarkers for Classifying Unipolar and Bipolar Depression Severity Yassine Ouzar, Clémence Nineuil, Fouad Boutaleb, Emery Pierson, Ali Amad, and Mohamed Daoudi Univ. Lille, CNRS, Centrale Lille, Institut Mines-Télécom, UMR 9189 CRIStAL, F-59000 Lille, France Univ. Lille, Inserm, CHU Lille, U1172 - LilNCog - Lille Neuroscience & Cognition, F-59000 Lille, France LIX, École Polytechnique, IPP Paris IMT Nord Europe, Institut Mines-Télécom, Univ. Lille, Centre for Digital Systems, F-59000 Lille, France Abstract Depression is a complex mental disorder characterized by a diverse range of observable and measurable indicators that go beyond traditional subjective assessments. Recent research has increasingly focused on objective, passive, and continuous monitoring using wearable devices to gain more precise insights into the physiological and behavioral aspects of depression. However, most existing studies primarily distinguish between healthy and depressed individuals, adopting a binary classification that fails to capture the heterogeneity of depressive disorders. In this study, we leverage wearable devices to predict depression subtypes—specifically unipolar and bipolar depression—aiming to identify distinctive biomarkers that could enhance diagnostic precision and support personalized treatment strategies. To this end, we introduce the CALYPSO dataset, designed for non-invasive detection of depression subtypes and symptomatology through physiological and behavioral signals, including blood volume pulse, electrodermal activity, body temperature, and three-axis acceleration. Additionally, we establish a benchmark on the dataset using well-known features and standard machine learning methods. Preliminary results indicate that features related to physical activity, extracted from accelerometer data, are the most effective in distinguishing between unipolar and bipolar depression, achieving an accuracy of 96.77%. Temperature-based features also showed high discriminative power, reaching an accuracy of 93.55%. These findings highlight the potential of physiological and behavioral monitoring for improving the classification of depressive subtypes, paving the way for more tailored clinical interventions. Fig. 1: Overview of the proposed framework for classifying depression types based on physiological signals collected from the Empatica E4 wristband. The initial step involves pre-processing the raw sensor data to remove noise and artifacts, thereby enhancing the signal-to-noise ratio. Subsequently, time-domain, frequency-domain, and statistical features are extracted from the filtered signals. Finally, these extracted features are fed individually and combined to ML algorithms for the classification of unipolar and bipolar depression. I INTRODUCTION Depression is a common psychiatric disorders affecting approximately 280 million people worldwide, accounting for 3.8% of the global population. Its prevalence is higher among women and individuals over the age of 60 [ 20 ] . The main symptoms of depression include persistent feelings of sadness, a loss of interest or pleasure in daily activities, sleep disturbances such as insomnia, an inability to experience joy, and, in severe cases, suicidal thoughts [ 14 ] . Although cultural, psychological, and biological factors are recognized as contributing to depression, and effective treatments such as psychological therapies and medications exist, the complex underlying pathogenesis remains poorly understood [ 14 ] . A significant barrier to progress lies in the reliance on subjective assessment tools, such as self-reported questionnaires and clinician-administered scales. These methods are prone to biases and the retrospective nature of the information collected [ 32 ] . To address these limitations, there is growing interest in objective assessment methods, particularly those utilizing wearable devices [ 26 , 24 ] . Wearable technology offers several advantages for depression assessment and monitoring, including continuous real-time data collection, improved ecological validity, and the ability to measure objective biomarkers. On the other hand, prior research have focused mainly on discriminate between healthy and depressed individuals. However, depression is a heterogeneous condition characterized by diverse clinical subtypes and categories such as unipolar and bipolar disorder, each with distinct behavioral, physiological, and neurobiological signatures [ 7 ] . Unipolar depression, often referred to as major depressive disorder, is characterized by persistent depressed mood, loss of interest, recurrent thoughts of death, and a range of physical and cognitive symptoms, typically occurring without episodes of mania [ 16 ] . In contrast, bipolar depression shares many overlapping symptoms with unipolar depression but is distinguished by the presence of alternating episodes of depression and mania or hypomania [ 18 ] . Bipolar patients may also exhibit atypical features, such as hypersomnia, hyperphagia, and psychotic features, which are less common in unipolar depression [ 19 ] . This distinction highlights the complexity and heterogeneity of depressive disorders. Therefore, simply distinguishing between healthy and depressed overlooks the important variations within depression that can significantly impact clinical outcomes and treatment approaches. In light of the limitations of previous studies, particularly their reliance on subjective assessments and limited exploration of depression subtypes, we aimed to investigate the potential of objective markers derived from wearable device in discriminating between unipolar and bipolar depression. This study is, to our knowledge, the first to specifically focus on detecting depression subtypes, moving beyond the typical binary distinction of healthy vs. depressed individuals. The main contributions of this study are summarized as follows: 1. We introduce a novel dataset named CALYPSO depression, specifically designed to detect various types of depression and their characteristics. Unlike most existing datasets, CALYPSO focuses on objective markers extracted from wearable devices, including HRV, EDA, temperature, and accelerometer data, alongside behavioral signals such as vocal tone, facial expressiveness, and head movements. 2. We establish a benchmark for differentiating between unipolar and bipolar depression using a comprehensive set of time-domain, frequency-domain, and statistical features derived from Empatica wearable sensor data. 3. We evaluated the extracted features, both individually and in combination, using several commonly used machine learning algorithms. II RELATED WORK By surveying existing research on depression, two primary diagnostic approaches can be identified, each differing in methodology and focus. Traditional approaches heavily rely on subjective assessments, such as patient self-reports, structured interviews, and clinician observations guided by DSM-5 criteria [ 3 ] . Although commonly used in clinical practice, this method has inherent limitations, including susceptibility to biases, inconsistencies in patient reporting, and reliance on retrospective information [ 32 ] . Moreover, the heterogeneous nature of depression poses significant challenges for accurate diagnosis [ 12 ] . In contrast, the second approach highlights the need for more objective and comprehensive diagnostic approaches that go beyond subjective assessments, integrating quantifiable biomarkers spanning behavioral, physiological, and neurobiological domains [ 24 ] . These include alterations in physiological responses, reduced physical activity, changes in vocal characteristics such as tone and prosody, and decreased facial expressiveness and emotional reactivity, among other indicators. Behavioral markers such as facial expressions, head movements, and vocal prosody are valuable indicators of affective states, including depression [ 11 ] . However, their reliability can be influenced by various factors, such as social and cultural differences, situational context, and environmental conditions (e.g., camera angles, lighting, ambiguity, and noise). Additionally, individuals can consciously control or mask their expressions, further complicating the interpretation of these markers [ 21 ] . On the other hand, physiological signals stand out as particularly important and promising. Signals such as heart rate variability (HRV), electrodermal activity (EDA), and respiratory patterns offer several key advantages over other markers. They directly reflect autonomic nervous system (ANS) activity, which is often dysregulated in individuals experiencing stress or emotional disturbances linked to depression [ 31 ] . Unlike subjective assessments or behavioral markers, physiological signals provide objective, quantifiable measurements that are less susceptible to subjective interpretation or patient bias [ 21 ] . Additionally, they enable passive, continuous, and real-time monitoring via wearable devices, enhancing ecological validity and reducing reliance on retrospective self-reports [ 6 ] . Wearable devices like wristbands and smartwatches are increasingly used to monitor physiological signals such as heart rate and skin conductance for depression detection [ 13 , 26 , 32 ] . Most existing studies have examined the use of these physiological cues separately. Hartmann et al. found significant correlations between depression and various heart rate variability features. Specifically, individuals with major depression tended to exhibit reduced high-frequency HRV (HRV-HF), low-frequency HRV (HRV-LF), very low-frequency HRV (HRV-VLF), standard deviation of normal-to-normal intervals (SDNN), and root mean square of successive differences (RMSSD), along with an increased LF/HF ratio when compared to healthy individuals [ 10 ] . Regarding electrodermal activity, studies have consistently shown that patients with depression exhibit lower skin conductance levels during rest compared to healthy control subjects [ 24 , 12 ] . Physical activity has also been linked to depression. A meta-analysis by Schuch et al. reported that greater physical activity, as measured by accelerometers, was associated with lower rates of depression [ 28 ] . Additionally, research on body temperature in depression has yielded interesting results. Tazawa et al. demonstrated that individuals with depression may exhibit elevated skin temperatures compared to healthy individuals [ 30 ] . While these studies have significantly advanced the understanding of physiological distinctions between healthy individuals and those with depression, they predominantly focus on binary classification (i.e., depressed vs. healthy) [ 12 ] . To the best of our knowledge, no existing research has explored the potential of wearable-derived physiological and behavioral data to discriminate between categories and subtypes of depression. Identifying physiological markers that distinguish between depression subtypes, such as unipolar, bipolar, melancholic, atypical, or psychotic depression, could provide deeper insights into the underlying mechanisms and support personalized treatment approaches. III MATERIALS AND METHODS III-A Dataset The data presented herein are derived from the CALYPSO clinical trial. This monocentric pilot study is designed to identify and describe objective markers associated with the severity of depression and its clinical characteristics. The CALYPSO clinical trial has been reviewed and approved by the regional ethics Committee, ensuring adherence to ethical standards and guidelines. All patients provided written informed consent before participation in the study. In routine psychiatric practice, patients admitted to the CHU of Lille undergo routine diagnostic evaluations conducted by an attending psychiatrist. A standardized clinical evaluation is performed, including a collection of medical history, a structured interview using the Hamilton Depression Rating Scale to assess the severity of depression, and an investigation of the specific characteristics of severe depression according to the DSM-5 criteria (melancholic, catatonic, psychotic, anxious, or seasonal) [ 3 ] . Patients who meet the DSM-5 inclusion criteria for severe depression are invited to participate in a more comprehensive interview, contributing directly to the CALYPSO depression dataset. A total of 40 eligible patients diagnosed with severe depression were included in the study. For each participant, an interview was recorded using a camera to capture facial expressions and body movements, and a microphone to analyze vocal features. Additionally, an Empatica E4 wristband was used to monitor physical activity and physiological responses. This study focuses specifically on the signals collected from the Empatica. The participants were predominantly French nationals (ethnicity data were not collected). Each patient took part in a structured clinical interview conducted in a controlled environment, as illustrated in Figure 2 . The interview consisted of two phases: an initial informal segment involving a brief, non-medical conversation with the clinician, followed by a structured assessment based on the Hamilton Depression Rating Scale (HDRS) to evaluate the severity of depression. On average, the maximum duration of the interviews was 20.53 seconds, with a standard deviation of 8.68 seconds. After the session, the clinician determined the depression subtype and its characteristics based on the patient’s responses and observed behaviors. The E4 wristband records six physiological signals in CSV format, including blood volume pulse (BVP), measured using photoplethysmography, average heart rate (HR) computed over 10-second intervals, and interbeat intervals (IBI), which measure the time between heartbeats and provide insights into heart rate variability. It also captures electrodermal activity (EDA), which reflects emotional responses through changes in skin conductance, accelerometer data (ACC), which monitors physical movement and activity, and skin temperature (TEMP), which indicates ANS function. The performance of the E4 bracelet has been validated in multiple studies [ 17 , 5 ] , and has been applied in diverse research fields, including sleep monitoring [ 29 ] , driving safety [ 2 ] , and social stress monitoring [ 27 ] . Data from 31 participants were available for analysis, as data from 9 participants were excluded due to signal corruption or missing information. Among the six physiological signals recorded by the Empatica E4, only four were used in this study, namely BVP, EDA, ACC, and TEMP. The other two signals were excluded due to the presence of certain corruptions in some recordings and because they are not directly relevant for quantifying depression. It is worth noting that IBI represents the time between successive systolic peaks in the BVP signals, and HR is calculated as 60/IBI to obtain the heart rate in beats per minute [ 23 ] . Fig. 2: Interview Room Setup for the Calypso Depression Dataset. III-B Features extraction We describe the set of interpretable features used in this paper. The extracted features include time-domain and frequency-domain heart rate variability (HRV) features derived from BVP signals, alongside features from EDA, ACC, and TEMP data. A detailed description of each feature category is provided below. III-B 1 Heart rate variability features Heart rate variability is a valuable indicator for analyzing the physiological mechanisms associated with depression. It reflects the balance between the activity of the sympathetic nervous system, linked to stress responses, and the parasympathetic nervous system, associated with relaxation and recovery [ 25 ] . Research has shown that low HRV is strongly correlated with the presence of depressive disorders [ 10 , 4 , 9 ] . Two types of HRV features were extracted from BVP signals in time-domain and frequency-domain. Following the pre-processing procedure described in [ 22 ] , detrending and a 2nd-order Butterworth band-pass filter were performed to eliminate noise and irrelevant frequency bands while preserving the pulse waveform information. Then, 25 time-domain and 10 frequency- domain HRV features were extracted from the filtered BVP signals using NeuroKit2 toolkit [ 15 ] . Details on both are given below. Time-domain heart rate variability features HRV Features Description MeanNN Mean of Normal-to-Normal (NN) Intervals SDNN Standard Deviation of NN Intervals SDANN1, SDANN2, SDANN5 Standard Deviation of Averages of NN Intervals over 1, 2, & 5 minutes SDNNI1, SDNNI2, SDNNI5 Standard Deviation of NN Intervals within segments of 1, 2, & 5 minutes RMSSD Root Mean Square of Successive Differences between adjacent NN intervals SDRMSSD Standard Deviation of RMSSD SDSD Standard Deviation of Successive Differences between adjacent NN intervals CVNN Coefficient of Variation of NN Intervals MCVNN Mean Coefficient of Variation of NN Intervals between adjacent NN intervals CVSD Coefficient of Variation of Successive Differences between adjacent NN intervals IQRNN Interquartile Range of NN Intervals MinNN & MaxNN Min and Max of NN Intervals MedianNN Median of NN Intervals MADNN Median Absolute Deviation of NN Intervals HTI Heart Rate Turbulence Index TINN Triangular Index of NN Intervals pNN50 & pNN20 Percentage of Pairs of Successive NN Intervals Prc20NN & Prc80NN 20th and 80th Percentille of NN Intervals TABLE I: Time-domain HRV features and their description. Time-domain HRV features primarily reflect the beat-to-beat variations in the time intervals between successive heartbeats, known as the IBI or NN (normal-to-normal) intervals. These variations mirror the overall ANS regulation, particularly the balance between the parasympathetic and sympathetic branches. These features provide insights into how the body responds to stress, recovery, and emotional regulation. 25 time-domain HRV features are returned by NeuroKit2 which are described in Table I . However, SDANN5 and SDNNI5, calculated from the variability of IBI across 5-minute segments, were excluded due to missing values (NaN) in some recordings. Frequency-domain heart rate variability features Frequency-domain HRV features analyze the periodic oscillations in heart rate at different frequencies, providing insights into the ANS regulation. These features are derived by analyzing the power spectral density of the IBI time series, which decomposes the HRV signal into distinct frequency bands, each associated with different physiological processes. The main frequency bands are the very high frequency (VHF: \geq 0.4 Hz), high frequency (HF: 0.15-0.4 Hz), low frequency (LF: 0.04-0.15 Hz), very low frequency (VLF: 0.0033-0.04 Hz), and ultra-low frequency (ULF: \leq 0.003 Hz). HRV features in the frequency domain are computed as the area under the PSD curve within the corresponding frequency band. For example, LF and HF powers are computed as the area under the PSD curve corresponding to 0.04-0.15Hz and 0.15-0.4Hz respectively (see Figure 3 ). NeuroKit2 returns 10 frequency-domain HRV features, which are detailed in Table II . Similar to time-domain HRV features, the ULF was excluded due to the presence of missing values. ULF power is typically measured over a long-term segment, often ranging from 24 hours to several hours of continuous IBI data. The time segment should be sufficiently long to capture the very low-frequency oscillations (below 0.0033 Hz), which occur over longer periods, making shorter recordings inadequate for ULF measurement. HRV Features Description TP Total variance of the HRV signal, reflecting overall autonomic activity ULF Power in the ultra low-frequency range (less than 0.0033 Hz), often associated with long-term regulatory mechanisms VLF Power in the very low-frequency range (0.0033 to 0.04 Hz), reflecting long-term regulatory mechanisms LF Power in the low-frequency range (0.04 to 0.15 Hz), associated with both sympathetic and parasympathetic activity HF Power in the high-frequency range (0.15 to 0.4 Hz), primarily reflecting parasympathetic activity (vagal). VHF Power in the very high-frequency range (above 0.4 Hz), often linked to respiration and higher autonomic modulation LF/HF Ratio The ratio of LF to HF, indicating the balance between sympathetic and parasympathetic activity LFn Normalized LF, calculated by dividing LF by TP HFn Normalized HF, calculated by dividing HF by TP LnHF The natural logarithm of HF, often used to normalize the feature and improve its interpretability TABLE II: Frequency-domain HRV features and their description. Fig. 3: A representative PSD for IBI signal showing The areas of VLF, LF, HF and VHF powers of the HRV. III-B 2 Electrodermal activity features Electrodermal activity (EDA) has been widely used as indicators of emotional and mental health states, including depression. EDA measures the variations in the electrical properties of the skin, resulting from the activity of sweat glands, recorded as changes in potential or variations in skin resistance. During emotional arousal, increased cognitive load, or physical activity, sweat production increases, leading to changes in the skin’s properties by raising conductivity and lowering resistance. Even a small amount of sweat, invisible to the naked eye on the skin’s surface, can cause a noticeable change in its electrical conductivity [ 21 ] . EDA signals were decomposed into two distinct components, namely phasic and tonic (see Figure 4 ). The tonic component represents the slow, continuous variations in electrodermal activity, capturing the baseline or background level of physiological arousal, such as chronic stress or anxiety. This component is extracted by isolating the low-frequency baseline variations in the raw signal. Once the tonic component is identified, it is subtracted from the raw EDA signal to obtain the phasic component. The phasic signal reflects the rapid and transient fluctuations in electrodermal activity, typically occurring in response to specific stimuli. These brief changes are linked to momentary activations of the sympathetic nervous system and are used to study stimulus-driven emotional or cognitive reactions. From each component, mean, standard deviation, minimum and maximum were computed as EDA features as well as the SCR amplitude of the signal excluding the tonic component ( SCR_{\text{Amplitude}} ) and the samples at which the onsets of the peaks occur ( SCR_{\text{Onsets}} ). Fig. 4: Visualization of an EDA signal. The top panel shows the raw and cleaned EDA signal. The middle panel displays the skin conductance response (SCR), highlighting the phasic component with identified SCR-onsets, SCR-peaks, and SCR-half recovery points. The bottom panel represents the skin conductance level (SCL), highlighting the tonic component of the EDA signal over time. III-B 3 Physical activity features Numerous cross-sectional studies have demonstrated a strong association between depression and increased sedentary behavior, likely attributed to reduced motivation and energy levels [ 1 ] . However, the relationship between physical activity and depression is complex and can vary among individuals. To quantify physical activity, accelerometer data (ACC) is collected using the Empatica E4 wristband, which is equipped with a 3-axis accelerometer. This device captures movement in three dimensions (x, y, z) at a sampling frequency of 32 Hz, enabling the detection of activity patterns such as walking, running, resting, and even subtle motion characteristics. Building upon the pre-processing steps performed in [ 8 ] , the ACC data was processed using a 5th-order Butterworth low-pass filter with a cutoff frequency of 10 Hz. Subsequently, the filtered ACC data was transformed into motion features by computing the total accelerometer magnitude ( ACC_{Magnitude} ) which is the vector norm of the three-axis x, y and z using the following formula: ACC_{Magnitude}=\sqrt{ACC_{x}^{2}+ACC_{y}^{2}+ACC_{z}^{2}} where ACC_{x} , ACC_{y} , ACC_{z} represent the accelerometer components in the x, y and z axes, respectively. Key features extracted from accelerometer magnitude include measures of movement intensity (mean( ACC_{Mean} ), max( ACC_{Max} ), min( ACC_{Min} ), standard deviation ( ACC_{STD} ), and energy( ACC_{Energy} )), movement dynamics (dominant frequency ( ACC_{Dominant-frequency} ), inactivity time( ACC_{Inactivity-time} )), and movement symmetry (correlations between accelerometer signals on different axes ( Symmetry_{(}x-y) , Symmetry_{(}y-z) , Symmetry_{(}x-z) )). The movement intensity measures ( ACC_{Mean} , ACC_{Max} , ACC_{Min} , ACC_{STD} ) are derived directly from the ACC_{Magnitude} and correspond to its average, maximum, minimum, and standard deviation, respectively. ACC_{Energy} is defined as the sum of the squared magnitudes normalized by the ACC signal length : ACC_{\text{Energy}}=\frac{\sum_{i=1}^{N}ACC_{Magnitude_{i}}^{2}}{N} where N is the total number of samples. This feature quantifies the overall power of the accelerometer magnitude signal. ACC_{Dominant\_frequency} reflects the most significant periodic motion in the signal, often associated with repetitive activities such as walking or running. It is calculated using Fast Fourier Transform (FFT) and represents the frequency with the highest amplitude in the power spectrum. It is expressed as follows: ACC_{\text{Dominant\_frequency}}=argmax(|FFT(ACC_{\text{Magnitude}})|)\times\frac{sampling\_rate}{length(ACC_{\text{Magnitude}})} where FFT is the result of applying the FFT to the accelerometer magnitude signal. argmax identifies the index of the maximum value in the magnitude of the FFT result. sampling\_rate is the sampling frequency of the ACC data. length(ACC_{\text{Magnitude}}) is the length of the signal used in the FFT computation. ACC_{Inactivity_{t}ime} quantifies periods with minimal or no movement, potentially indicating sedentary behavior or sleep. It is calculated using the following formula: ACC_{\text{Inactivity\_time}}=\frac{\sum_{i=1}^{N}(ACC_{Magnitude_{i}}