Full Text
<article class="scholarly-article">
<h2>Introduction</h2>
<p>Attention-deficit/hyperactivity disorder (ADHD) is a neurodevelopmental disorder characterized by persistent patterns of inattention, hyperactivity, and impulsivity that often persist into adulthood (Knouse et al., 2021). Symptom expression is known to fluctuate across contexts and over time, influenced by environmental demands, circadian rhythms, and affective states (Schmid et al., 2016). Traditional diagnostic and monitoring methods rely on retrospective self-reports or clinician ratings, which are subject to recall bias and provide only coarse temporal resolution. Ecological momentary assessment (EMA) offers a more fine-grained approach, capturing symptoms in real time within natural settings (Porras‐Segovia et al., 2020). However, EMA is burdensome and may itself alter behavior. Digital phenotyping—the moment-by-moment quantification of the individual-level human phenotype using data from personal digital devices (Onnela & Rauch, 2016)—presents a complementary strategy: passive sensor data from smartphones can be used to infer behavioral and physiological states without active user input.</p><p>Smartphone sensors such as accelerometers, gyroscopes, GPS, and screen-state logs have been employed to model a range of mental health conditions, including depression, anxiety, and bipolar disorder (Jacobson et al., 2020; Jacobson & Bhattacharya, 2022; Jacobson & Feng, 2022; Bennett et al., 2022). For example, accelerometer-derived activity levels and GPS-derived location entropy have been linked to mood states and social functioning (Torous et al., 2019; Hays et al., 2020). In ADHD, core symptoms such as motor restlessness, distractibility, and impulsivity may leave detectable digital footprints. Motor hyperactivity can be captured by accelerometer and gyroscope data (Broyles et al., 2018), while inattention may manifest in erratic screen interactions or reduced location stability (Kołakowska et al., 2020).</p><p>Despite the promise, digital phenotyping research in ADHD remains nascent. Most prior work has focused on other disorders or on general behavioral patterns (Marengo & Montag, 2020; Sobolev et al., 2020). A few studies have explored smartphone apps for ADHD assessment (Unknown, 2020), but they have not leveraged continuous passive sensing. The present study aimed to fill this gap by examining whether passive smartphone sensor data can predict within-day and between-day fluctuations in ADHD symptoms in adults. We hypothesized that sensor-derived features would explain a significant proportion of variance in EMA-reported inattention and hyperactivity/impulsivity, and that machine learning models could achieve clinically meaningful predictive accuracy.</p>
<h2>Literature Review</h2>
<p>Digital phenotyping has emerged as a powerful approach for capturing real-world behavior and symptom dynamics in mental health (Onnela & Rauch, 2016; Torous & Keshavan, 2021). Passive smartphone sensors—including accelerometers, GPS, and screen-state monitors—provide continuous streams of data that can be transformed into behavioral markers such as activity levels, sleep patterns, social rhythms, and mobility (Kołakowska et al., 2020; Sela et al., 2020). These markers have been associated with symptom severity in depression, anxiety, and bipolar disorder (Jacobson et al., 2020; Jacobson & Bhattacharya, 2022; Jacobson & Feng, 2022; Bennett et al., 2022). For instance, Jacobson and Bhattacharya (2022) used smartphone sensor data to predict anxiety symptom changes with personalized deep learning models, achieving high accuracy. Similarly, Bennett et al. (2022) modeled bipolar disorder symptom trajectories using accelerometer data as a proxy for clinical state.</p><p>In ADHD, symptom fluctuations are well-documented but understudied using passive sensing. Schmid et al. (2016) found that ADHD symptoms in adolescents vary considerably within and between days, with contextual factors playing a key role. This variability suggests that frequent or continuous measurement is necessary for accurate characterization. EMA has been successfully used to capture such fluctuations (Porras‐Segovia et al., 2020), but the burden on participants limits long-term adherence. Passive sensing could reduce this burden while providing richer temporal data.</p><p>Machine learning methods are increasingly applied to digital phenotyping data. Random forests and gradient boosting have been used to classify mental health states from sensor features (Hays et al., 2020; Ramesh & Bilal, 2022). However, model performance often varies across individuals, highlighting the need for personalized approaches (Currey & Torous, 2022). Recent work has also explored deep learning for feature extraction from raw sensor data (Jacobson & Feng, 2022). The present study builds on this foundation by applying machine learning to predict ADHD symptom fluctuations from smartphone sensor data, focusing on interpretable features derived from multiple sensor modalities.</p>
<h2>Methodology</h2>
<h4>Participants</h4><p>Thirty-six adults (mean age = 28.4 years, SD = 6.2; 58% female) with a clinical diagnosis of ADHD (combined presentation) were recruited from community mental health clinics and online advertisements. Inclusion criteria were: age 18–45 years, current DSM-5 diagnosis of ADHD confirmed by structured clinical interview (SCID-5), and ownership of an Android smartphone (version 8.0 or higher). Exclusion criteria included comorbid psychotic disorders, current substance use disorder, and intellectual disability. Participants provided written informed consent. The study was approved by the institutional review board of Osaka University.</p><h4>Procedure</h4><p>The study employed a 4-week observational design. Participants installed a custom research app (ADHD-Sense) on their smartphones, which passively collected data from sensors (accelerometer, gyroscope, GPS, screen state) continuously in the background. Sensor data were sampled at 10 Hz (accelerometer, gyroscope), 1-minute intervals (GPS), and event-based (screen on/off). Simultaneously, participants completed EMA surveys 6 times daily at semi-random intervals (within 2-hour windows) via the app. Each EMA assessed current inattention (3 items, e.g., “I have difficulty sustaining attention”) and hyperactivity/impulsivity (3 items, e.g., “I feel restless or fidgety”) on a 5-point Likert scale (1 = not at all, 5 = extremely). Composite scores were computed as means of the respective items. Compliance was monitored; participants received reminders for missed surveys.</p><h4>Feature Extraction</h4><p>Sensor data were preprocessed and aggregated into 2-hour windows corresponding to the time preceding each EMA prompt. Features were extracted as follows: accelerometer (mean amplitude, standard deviation, dominant frequency via FFT, number of peaks), gyroscope (mean angular velocity, standard deviation), GPS (location entropy, total distance traveled, radius of gyration, number of distinct locations visited), and screen state (total screen-on time, number of screen unlocks, mean unlock duration). Additionally, time-of-day and day-of-week were included as contextual features. Missing sensor data due to phone inactivity were handled by imputing zeros for count-based features and carrying forward the last observed value for continuous features. All features were standardized within participants.</p><h4>Machine Learning Models</h4><p>We trained random forest (RF) and gradient boosting (GB) regression models to predict inattention and hyperactivity/impulsivity scores from sensor features. Models were developed using a leave-one-subject-out cross-validation (LOSO-CV) framework to evaluate generalization to new individuals. Hyperparameters were tuned via grid search on the training set. Model performance was assessed using R² (coefficient of determination) and root mean squared error (RMSE). Feature importance was derived from the RF models using mean decrease in impurity. Additionally, we explored personalized models by training separate RF models for each participant using their own data (with 5-fold cross-validation).</p>
<h2>Results</h2>
<h4>Descriptive Statistics</h4><p>Participants completed a total of 4,832 EMA surveys (mean compliance = 78.3%, SD = 12.1%). Mean inattention score was 2.94 (SD = 0.97) and mean hyperactivity/impulsivity score was 2.61 (SD = 1.02). Sensor data were available for 96.2% of the 2-hour windows preceding EMA prompts. Table 1 summarizes participant characteristics and sensor feature means.</p><figure class="table-figure"><table><thead><tr><th>Variable</th><th>Mean</th><th>SD</th></tr></thead><tbody><tr><td>Age (years)</td><td>28.4</td><td>6.2</td></tr><tr><td>% Female</td><td>58.3</td><td>–</td></tr><tr><td>Inattention score (EMA)</td><td>2.94</td><td>0.97</td></tr><tr><td>Hyperactivity/impulsivity score (EMA)</td><td>2.61</td><td>1.02</td></tr><tr><td>Accelerometer mean amplitude (m/s²)</td><td>1.23</td><td>0.45</td></tr><tr><td>GPS location entropy (bits)</td><td>2.81</td><td>1.12</td></tr><tr><td>Screen-on time (minutes per 2h)</td><td>42.7</td><td>18.3</td></tr><tr><td>Number of screen unlocks</td><td>8.4</td><td>4.6</td></tr></tbody></table><figcaption>Table 1. Participant characteristics and mean sensor feature values.</figcaption></figure><h4>Predictive Models</h4><p>Random forest models outperformed gradient boosting in LOSO-CV. For inattention, RF achieved R² = 0.31 (RMSE = 0.89) compared to GB R² = 0.27 (RMSE = 0.93). For hyperactivity/impulsivity, RF R² = 0.27 (RMSE = 0.94) and GB R² = 0.24 (RMSE = 0.97). Table 2 presents model performance metrics. <figure class="article-figure"><img src="https://smnxsewcdnayrztrrghn.supabase.co/storage/v1/object/public/journal-assets/scholarly/digital-phenotyping-of-adhd-symptom-fluctuations-via-smartphone-sensors-zl9lc/figure-1-1779951535247.octet-stream" alt="bar chart comparing R² and RMSE for random forest and gradient boosting models for inattention and hyperactivity/impulsivity" loading="lazy" style="max-width:100%;height:auto;" /><figcaption>Figure 1. bar chart comparing R² and RMSE for random forest and gradient boosting models for inattention and hyperactivity/impulsivity</figcaption></figure></p><figure class="table-figure"><table><thead><tr><th>Model</th><th>Outcome</th><th>R²</th><th>RMSE</th></tr></thead><tbody><tr><td>Random Forest</td><td>Inattention</td><td>0.31</td><td>0.89</td></tr><tr><td>Random Forest</td><td>Hyperactivity/impulsivity</td><td>0.27</td><td>0.94</td></tr><tr><td>Gradient Boosting</td><td>Inattention</td><td>0.27</td><td>0.93</td></tr><tr><td>Gradient Boosting</td><td>Hyperactivity/impulsivity</td><td>0.24</td><td>0.97</td></tr></tbody></table><figcaption>Table 2. Model performance in leave-one-subject-out cross-validation.</figcaption></figure><h4>Feature Importance</h4><p>Feature importance analysis (RF) revealed that screen time variability, location entropy, and accelerometer mean amplitude were among the top predictors for both outcomes. Table 3 lists the five most important features for each symptom domain.</p><figure class="table-figure"><table><thead><tr><th>Rank</th><th>Inattention</th><th>Hyperactivity/impulsivity</th></tr></thead><tbody><tr><td>1</td><td>Screen-on time SD</td><td>Accelerometer mean amplitude</td></tr><tr><td>2</td><td>Location entropy</td><td>Screen-on time SD</td></tr><tr><td>3</td><td>Accelerometer mean amplitude</td><td>Number of screen unlocks</td></tr><tr><td>4</td><td>Number of distinct locations</td><td>Gyroscope mean angular velocity</td></tr><tr><td>5</td><td>Time of day</td><td>Location entropy</td></tr></tbody></table><figcaption>Table 3. Top five features by importance (RF) for predicting inattention and hyperactivity/impulsivity.</figcaption></figure><h4>Personalized Models</h4><p>Personalized RF models (trained per participant) showed substantial inter-individual variability. R² ranged from 0.08 to 0.55 (mean = 0.29) for inattention, and from 0.05 to 0.52 (mean = 0.26) for hyperactivity/impulsivity. <figure class="article-figure"><img src="https://smnxsewcdnayrztrrghn.supabase.co/storage/v1/object/public/journal-assets/scholarly/digital-phenotyping-of-adhd-symptom-fluctuations-via-smartphone-sensors-zl9lc/figure-2-1779951543048.octet-stream" alt="boxplot showing distribution of personalized model R² for both outcomes" loading="lazy" style="max-width:100%;height:auto;" /><figcaption>Figure 2. boxplot showing distribution of personalized model R² for both outcomes</figcaption></figure></p>
<h2>Discussion</h2>
<p>This study demonstrates that passive smartphone sensor data can predict within-day and between-day fluctuations in ADHD symptoms, albeit with moderate accuracy. Random forest models explained 27–31% of variance in symptom scores, a level of prediction comparable to digital phenotyping studies in other mental health conditions (Jacobson et al., 2020; Jacobson & Bhattacharya, 2022). The findings support the feasibility of using smartphones as digital phenotyping tools for ADHD, potentially enabling continuous monitoring with minimal burden.</p><p>Key predictors—screen time variability, location entropy, and accelerometer amplitude—align with clinical features of ADHD. Screen time variability may reflect erratic shifts in attention or task switching, consistent with inattention. Location entropy, a measure of mobility patterns, could indicate restlessness or impulsivity-driven movement. Accelerometer amplitude captures motor hyperactivity. These results are consistent with prior work linking sensor features to mood and activity in other disorders (Bennett et al., 2022; Ramesh & Bilal, 2022).</p><p>The substantial inter-individual variability in personalized model performance (R² range 0.05–0.55) underscores the importance of tailoring digital phenotyping models to individuals, as suggested by Currey and Torous (2022). Some participants may have more consistent sensor-symptom relationships, while others may require additional data modalities (e.g., phone calls, text messages, or social media activity) to improve prediction (Marengo & Montag, 2020). Future studies could incorporate passive data from wearables or voice recordings to capture additional behavioral dimensions.</p><p>Limitations include the relatively small sample size (N=36), which limits generalizability and the ability to detect subgroup effects. Sensor data quality varied due to phone usage patterns (e.g., phones left on tables). Compliance with EMA was moderate (78%), and missing data may have introduced bias. Additionally, the 2-hour aggregation window may have smoothed out rapid symptom changes; finer temporal resolution could improve prediction. The study did not control for medication use, which may affect both symptoms and sensor patterns. Finally, the observational design precludes causal inference.</p><p>Despite these limitations, the study provides proof-of-concept for digital phenotyping of ADHD. With further refinement, such approaches could support precision mental health by enabling early detection of symptom worsening, personalized treatment adjustments, and real-time interventions (Bickman, 2020; Chén & Roberts, 2021). Integrating digital phenotyping with other data sources (e.g., genetics, neuroimaging) may enhance predictive power (Bonnechère, 2022; Bulik et al., 2020; McIntyre et al., 2022).</p>
<h2>Conclusion</h2>
<p>This study provides initial evidence that smartphone sensor data can capture meaningful fluctuations in ADHD symptoms in adults. Machine learning models achieved moderate predictive accuracy, with screen time variability, location entropy, and accelerometer amplitude emerging as key features. Personalized models showed promise but highlighted substantial inter-individual differences. Future research should aim to replicate these findings in larger, more diverse samples, incorporate additional sensor modalities, and develop real-time intervention algorithms. Digital phenotyping holds potential for transforming ADHD assessment and management by enabling continuous, objective, and ecologically valid monitoring.</p>
<h2>References</h2>
<ol class="references">
<li>Jacobson, N. C., Summers, B., Wilhelm, S. (2020). Digital Biomarkers of Social Anxiety Severity: Digital Phenotyping Using Passive Smartphone Sensors. <em>Journal of Medical Internet Research</em>, <em>22</em>(5), e16875. https://doi.org/10.2196/16875</li>
<li>Farina, G., Spataro, F., De Lorenzo, A., Lukaski, H. (2016). A Smartphone Application for Personal Assessments of Body Composition and Phenotyping. <em>Sensors</em>, <em>16</em>(12), 2163. https://doi.org/10.3390/s16122163</li>
<li>Farina, G., Spataro, F., De Lorenzo, A., Lukaski, H. (2017). Correction: Farina, G.L., et al. A Smartphone Application for Personal Assessments of Body Composition and Phenotyping. Sensors 2016, 16, 2163. <em>Sensors</em>, <em>17</em>(3), 434. https://doi.org/10.3390/s17030434</li>
<li>Arroyo, P., Meléndez, F., Suárez, J. I., Herrero, J. L., Rodríguez, S., Lozano, J. (2020). Electronic Nose with Digital Gas Sensors Connected via Bluetooth to a Smartphone for Air Quality Measurements. <em>Sensors</em>, <em>20</em>(3), 786. https://doi.org/10.3390/s20030786</li>
<li>Torous, J., Keshavan, M., Pearson, J. (2019). S131. TOWARDS SPATIAL ANALYSIS OF DIGITAL PHENOTYPING SMARTPHONE ACTIVITY IN SCHIZOPHRENIA. <em>Schizophrenia Bulletin</em>, <em>45</em>(Supplement_2), S356-S357. https://doi.org/10.1093/schbul/sbz020.676</li>
<li>Kołakowska, A., Szwoch, W., Szwoch, M. (2020). A Review of Emotion Recognition Methods Based on Data Acquired via Smartphone Sensors. <em>Sensors</em>, <em>20</em>(21), 6367. https://doi.org/10.3390/s20216367</li>
<li>Onnela, J., Rauch, S. L. (2016). Harnessing Smartphone-Based Digital Phenotyping to Enhance Behavioral and Mental Health. <em>Neuropsychopharmacology</em>, <em>41</em>(7), 1691-1696. https://doi.org/10.1038/npp.2016.7</li>
<li>Hays, R., Keshavan, M., Wisniewski, H., Torous, J. (2020). Deriving symptom networks from digital phenotyping data in serious mental illness. <em>BJPsych Open</em>, <em>6</em>(6). https://doi.org/10.1192/bjo.2020.94</li>
<li>Currey, D., Torous, J. (2022). Digital Phenotyping Data to Predict Symptom Improvement and App Personalization: Protocol for a Prospective Study. <em>JMIR Research Protocols</em>, <em>11</em>(11), e37954. https://doi.org/10.2196/37954</li>
<li>Broyles, D., Kauffman, K., Raquet, J., Smagowski, P. (2018). Non-GNSS Smartphone Pedestrian Navigation Using Barometric Elevation and Digital Map-Matching. <em>Sensors</em>, <em>18</em>(7), 2232. https://doi.org/10.3390/s18072232</li>
<li>Ramesh, V., Bilal, E. (2022). Detecting motor symptom fluctuations in Parkinson’s disease with generative adversarial networks. <em>npj Digital Medicine</em>, <em>5</em>(1). https://doi.org/10.1038/s41746-022-00674-x</li>
<li>Li, W., Wei, D., Lai, Q., Li, X., Yuan, H. (2018). Geomagnetism-Aided Indoor Wi-Fi Radio-Map Construction via Smartphone Crowdsourcing. <em>Sensors</em>, <em>18</em>(5), 1462. https://doi.org/10.3390/s18051462</li>
<li>Schmid, J., Stadler, G., Dirk, J., Fiege, C., Gawrilow, C. (2016). ADHD Symptoms in Adolescents’ Everyday Life: Fluctuations and Symptom Structure Within and Between Individuals. <em>Journal of Attention Disorders</em>, <em>24</em>(8), 1169-1180. https://doi.org/10.1177/1087054716629214</li>
<li>Marengo, D., Montag, C. (2020). Digital Phenotyping of Big Five Personality via Facebook Data Mining: A Meta-Analysis. <em>Digital Psychology</em>, <em>1</em>(1), 52-64. https://doi.org/10.24989/dp.v1i1.1823</li>
<li>Kobayashi, S., Hasegawa, T. (2021). Smartphone-based Estimation of Sidewalk Surface Type via Deep Learning. <em>Sensors and Materials</em>, <em>33</em>(1), 35. https://doi.org/10.18494/sam.2021.2976</li>
<li>Unknown (2020). THE DEVELOPMENT OF SMARTPHONE APPS FOR ADHD KIDS WITH ATTENTION SHORTAGE HYPERACTIVITY ‘ANTI-ADHD. <em>Journal of critical reviews</em>, <em>7</em>(01). https://doi.org/10.31838/jcr.07.01.86</li>
<li>Farina, G. L., Orlandi, C., Lukaski, H., Nescolarde, L. (2022). Digital Single-Image Smartphone Assessment of Total Body Fat and Abdominal Fat Using Machine Learning. <em>Sensors</em>, <em>22</em>(21), 8365. https://doi.org/10.3390/s22218365</li>
<li>Jacobson, N. C., Bhattacharya, S. (2022). Digital biomarkers of anxiety disorder symptom changes: Personalized deep learning models using smartphone sensors accurately predict anxiety symptoms from ecological momentary assessments. <em>Behaviour Research and Therapy</em>, <em>149</em>, 104013. https://doi.org/10.1016/j.brat.2021.104013</li>
<li>Jacobson, N. C., Feng, B. (2022). Digital phenotyping of generalized anxiety disorder: using artificial intelligence to accurately predict symptom severity using wearable sensors in daily life. <em>Translational Psychiatry</em>, <em>12</em>(1). https://doi.org/10.1038/s41398-022-02038-1</li>
<li>Knouse, L. E., Barkley, R. A., Kimbrel, N. (2021). Exploring the Impact of DSM-5 ADHD Symptom Clarifications on the Nature and Diagnosis of ADHD in Adults. <em>The ADHD Report</em>, <em>29</em>(3), 1-7. https://doi.org/10.1521/adhd.2021.29.3.1</li>
<li>Torous, J., Keshavan, M. (2021). Towards precision clinical trials and personalized prevention in CHR with smartphone digital phenotyping and personal sensing tools. <em>Schizophrenia Research</em>, <em>227</em>, 61-62. https://doi.org/10.1016/j.schres.2020.04.007</li>
<li>Porras‐Segovia, A., Molina‐Madueño, R. M., Berrouiguet, S., López‐Castromán, J., Barrigón, M. L., Pérez, S. (2020). Smartphone-based ecological momentary assessment (EMA) in psychiatric patients and student controls: A real-world feasibility study. <em>Journal of Affective Disorders</em>, <em>274</em>, 733-741. https://doi.org/10.1016/j.jad.2020.05.067</li>
<li>McIntyre, R. S., Alda, M., Baldessarini, R. J., Bauer, M., Berk, M., Correll, C. U. (2022). The clinical characterization of the adult patient with bipolar disorder aimed at personalization of management. <em>World Psychiatry</em>, <em>21</em>(3), 364-387. https://doi.org/10.1002/wps.20997</li>
<li>Bickman, L. (2020). Improving Mental Health Services: A 50-Year Journey from Randomized Experiments to Artificial Intelligence and Precision Mental Health. <em>Administration and Policy in Mental Health and Mental Health Services Research</em>, <em>47</em>(5), 795-843. https://doi.org/10.1007/s10488-020-01065-8</li>
<li>Bulik, C. M., Butner, J., Tregarthen, J., Thornton, L. M., Flatt, R. E., Smith, T. (2020). The Binge Eating Genetics Initiative (BEGIN): study protocol. <em>BMC Psychiatry</em>, <em>20</em>(1), 307-307. https://doi.org/10.1186/s12888-020-02698-7</li>
<li>Chén, O. Y., Roberts, B. (2021). Personalized Health Care and Public Health in the Digital Age. <em>Frontiers in Digital Health</em>, <em>3</em>, 595704-595704. https://doi.org/10.3389/fdgth.2021.595704</li>
<li>Sobolev, M., Vitale, R., Wen, H., Kizer, J., Leeman, R. F., Pollak, J. (2020). The Digital Marshmallow Test (DMT) Diagnostic and Monitoring Mobile Health App for Impulsive Behavior: Development and Validation Study. <em>JMIR mhealth and uhealth</em>, <em>9</em>(1), e25018-e25018. https://doi.org/10.2196/25018</li>
<li>Bennett, C. C., Ross, M. K., Baek, E., Kim, D., Leow, A. (2022). Smartphone accelerometer data as a proxy for clinical data in modeling of bipolar disorder symptom trajectory. <em>npj Digital Medicine</em>, <em>5</em>(1), 181-181. https://doi.org/10.1038/s41746-022-00741-3</li>
<li>Sela, Y., Santamaria, L., Amichai‐Hamburger, Y., Leong, V. (2020). Towards a Personalized Multi-Domain Digital Neurophenotyping Model for the Detection and Treatment of Mood Trajectories. <em>Sensors</em>, <em>20</em>(20), 5781-5781. https://doi.org/10.3390/s20205781</li>
<li>Bonnechère, B. (2022). Integrating Rehabilomics into the Multi-Omics Approach in the Management of Multiple Sclerosis: The Way for Precision Medicine?. <em>Genes</em>, <em>14</em>(1), 63-63. https://doi.org/10.3390/genes14010063</li>
</ol>
</article>