Source-linked AI summary
DeepMood: Modeling Mobile Phone Typing Dynamics for Mood Detection
Bokai Cao, Lei Zheng, Chenwei Zhang, Philip S. Yu, Andrea Piscitello, John Zulueta, Olu Ajilore, Kelly Ryan, Alex D. Leow
TL;DR
Prior work had not examined the relationship between mobile phone typing dynamics and mood states, a potential source of unobtrusive information about psychiatric illness in daily life. DeepMood uses late fusion to model session-level typing metadata and achieved 90.31% prediction accuracy on depression scores. The authors conclude that mobile phone metadata could help infer mood disturbance and severity.
Problem
The relationship between mobile phone typing dynamics and mood states had not previously been investigated, despite the potential to study psychiatric illness unobtrusively in daily life.
Source: §1 · p. 1Method
DeepMood separately models session-level time-series views and applies late fusion to their representations.
Source: §1 · pp. 1–2Results
90.31% prediction accuracy on the depression score was achieved using session-level mobile phone typing dynamics.
Source: §1 · p. 2Takeaways & Limitations
Mobile phone metadata could be used to infer the presence and severity of mood disturbances and may enable earlier intervention during mood episodes.
Source: §6 · p. 8Privacy concerns motivate restricting collection where the information could be used to recover text.
Source: §1 · p. 2
Abstract
from arXiv · showhide
The increasing use of electronic forms of communication presents new opportunities in the study of mental health, including the ability to investigate the manifestations of psychiatric diseases unobtrusively and in the setting of patients' daily lives. A pilot study to explore the possible connections between bipolar affective disorder and mobile phone usage was conducted. In this study, participants were provided a mobile phone to use as their primary phone. This phone was loaded with a custom keyboard that collected metadata consisting of keypress entry time and accelerometer movement. Individual character data with the exceptions of the backspace key and space bar were not collected due to privacy concerns. We propose an end-to-end deep architecture based on late fusion, named DeepMood, to model the multi-view metadata for the prediction of mood scores. Experimental results show that 90.31% prediction accuracy on the depression score can be achieved based on session-level mobile phone typing dynamics which is typically less than one minute. It demonstrates the feasibility of using mobile phone metadata to infer mood disturbance and severity.
1 INTRODUCTION
The introduction motivates unobtrusive, naturalistic monitoring and identifies typing dynamics as an unstudied link to mood states in bipolar disorder. DeepMood models session-level, multi-view typing data with late fusion; the paper reports 90.31% accuracy on depression scores.
1 INTRODUCTION
- Smartphones offer opportunities to study psychiatric illness unobtrusively in patients’ daily lives and at greater detail than previously possible.Automatically generated smartphone data in naturalistic settings could facilitate early intervention and serve as objective outcome measures in efficacy trials.Source: §1 · p. 1
- The authors identify typing dynamics and mood states in bipolar disorder as a previously unstudied relationship, despite text messaging’s frequent use.They describe bipolar disorder as costly and suggest early intervention and prevention management could reduce its financial impact.Source: §1 · p. 1
- Each session contains multiple views, such as alphanumeric, special-character, and accelerometer data, with different timestamps and densities.Sessions begin after at least five seconds without a keypress and end when another five-second gap occurs.Source: §1 · p. 1
- Timestamp alignment creates missing values across views, while concatenation risks allowing denser accelerometer data to overwhelm sparser typing metadata.The accelerometer stream has 16 times more data points than character-related metadata in the dataset; the authors also identify cross-view interactions as a modeling challenge.Source: §1 · p. 2
- DeepMood separately models each time-series view with an RNN, then fuses view outputs using Multi-view Machine, Factorization Machine, or fully connected approaches.The authors frame this two-stage late-fusion approach as a response to problems with aligning or concatenating views.Source: §1 · p. 2
- Privacy concerns motivate restricting collected information because it could otherwise be used to recover text.The passage links the restriction to the risk that a malicious person could unscramble and recover texts.Source: §1 · p. 2
- 90.31% prediction accuracy on the depression score was achieved using session-level mobile phone typing dynamics.The sessions are typically less than one minute long.Source: §1 · p. 2
2 DATA
This section characterizes the BiAffect cohort and privacy-limited phone metadata, then compares alphanumeric, special-key, and accelerometer patterns across mood states.
2 DATA
- 7 participants with bipolar I disorder, 5 with bipolar II, and 8 without diagnosis contributed data; the keyboard withheld character contents except backspace and space.HDRS and YMRS were administered weekly as symptom-rating standards, though these scales require face-to-face assessment.Source: §2 · pp. 2–3
2.1 Alphanumeric Characters
- Keypress-duration distributions differ by mood: mild depression trends shorter than controls, while manic samples tend to hold keys longer.Most keypresses are fast, but a non-negligible tail lasts longer.Source: Figure 3 · p. 3, §2.1 · p. 3
- Mood-disturbed samples show more restricted time-since-keypress distributions, while normal samples span a broader range of reactivity.The median interval is 380 ms, and distribution pairs differ significantly for 98.06% of HDRS and 99.52% of YMRS comparisons.Source: Figure 4 · p. 3, §2.1 · p. 3
- Distance from the previous key, a rough proxy for typing semantics, shows no clear distinction across mood states.The plotted lines are nearly identical.Source: Figure 5 · p. 3, §2.1 · p. 3
2.2 Special Characters
- Special-key rates show no obvious mood-state separation, though alphanumeric rate correlates negatively with backspace and switching-keyboard rate positively with other-key rate.Alphanumeric keys are generally most frequent per session, followed by autocorrect, space, and backspace.Source: §2.2 · pp. 3–4, Figure 6 · p. 4
2.3 Accelerometer Values
- Background accelerometer readings differ by mood, with disturbed samples showing larger Z-axis and smaller Y-axis values than normal samples.The authors hypothesize that this pattern reflects phone angle, given that collection occurred only in portrait position.Source: Figure 7 · p. 4, §2.3 · p. 4
3 DEEPMOOD ARCHITECTURE BASED ON LATE FUSION
DeepMood models each metadata view separately with recurrent units, then fuses their output representations to predict mood scores. Its late-fusion design avoids directly aligning heterogeneous time series and supports end-to-end training.
3 DEEPMOOD ARCHITECTURE BASED ON LATE FUSION
- DeepMood separately models each view with a GRU, then fuses the view outputs to predict mood scores.The fusion layers and GRUs are trained end to end by back-propagating the mood-score loss through the fusion layers to the recurrent parameters.Source: §3 · p. 5, §3.2 · p. 6
3.1 Modeling One View
- GRUs model sequential dependencies in each view, with reset and update gates that forget prior state and balance prior and candidate states.The paper motivates GRUs as a more effective choice for long sequences than traditional RNNs, which can suffer from exploding and vanishing gradients.Source: §3.1 · p. 5
3.2 Late Fusion on Multiple Views
- Late fusion combines view representations after separate modeling, avoiding direct early-fusion alignment and sampling-frequency problems.The paper identifies missing values and differences in view densities as challenges for early fusion.Source: §3.2 · p. 5
- Three fusion alternatives are considered: concatenation with fully connected layers, Factorization Machines for second-order interactions, and Multi-view Machines for interactions up to the number of views.The Multi-view Machine models full-order cross-view feature interactions in a collectively factorized tensor.Source: §3.2 · pp. 5–6
4 EXPERIMENTS
The experiments evaluate session-level mood prediction from alphanumeric, special-character, and accelerometer views, comparing late- and early-fusion methods. DeepMood’s late-fusion variants lead on dichotomized depression scores, while convergence and view analyses further examine model behavior.
4 EXPERIMENTS
- The session-level task uses alphanumeric-character, special-character, and accelerometer-value features from each session to predict its participant’s mood score.Sessions with more than 100 keypresses were truncated, and sessions with any view containing fewer than 10 keypresses were removed. The resulting 14,613 samples were split by time, with each user’s first 80% of sessions used for training and the rest for validation.Source: §4 · p. 6, §4.1 · p. 6
4.1 Experimental Setup
- HDRS is classified as normal at scores 0–7 and positive at scores of 8 or higher, whereas YMRS is treated as a regression target.Accuracy and F-score evaluate classification, and RMSE evaluates regression.Source: §4.1 · pp. 6–7
4.2 Compared Methods
- The comparison includes DeepMood with Multi-view Machine, Factorization Machine, and fully connected fusion, alongside XGB, SVM, and LR early-fusion baselines.Fusion-layer model complexity is controlled by setting the DNN fusion width to k′ = ck.Source: §4.2 · p. 7
4.3 Prediction Performance
- 90.31% (DMVM) and 90.21% (DFM) accuracy on dichotomized HDRS scores make the late-fusion DeepMood methods the best-performing compared methods.DMVM also outperforms XGB by 5.56% in accuracy, 5.93% in F-score, and 10.02% in RMSE.Source: Table 3 · p. 7, §4.3 · p. 7
4.4 Convergence Efficiency
- DMVM and DFM converge more efficiently than DNN during the first 300 epochs and reach a better final loss minimum; all DeepMood methods later outperform XGB.XGB's training error can reach zero early despite limited generalizability on sequence prediction.Source: Figure 10 · p. 8, §4.4 · p. 8
4.5 Importance of Different Views
- Alphanumeric features are the strongest individual mood predictors, accelerometer features also perform well, and special characters are poor predictors.The analysis links mood disturbance with typing patterns, including keypress duration and time since the previous keypress, as well as accelerometer values.Source: Table 4 · p. 8, §4.5 · p. 8
5 RELATED WORK
Related work covers sequence-prediction methods and multi-view learning. The paper positions complex event-sequence classification and deep modeling as areas needing further investigation.
5 RELATED WORK
- Sequence classification work is grouped into feature-based, distance-based, and model-based methods, with examples including conventional classifiers and probabilistic sequence models.The cited survey divides sequence data into five subtypes, including complex event sequences [50]; examples include feature-based methods, distance-based methods and, and model-based methods,, and.Source: §5 · p. 8
- Most prior work focuses on simple symbolic sequences or time series, while complex event-sequence classification and deep models remain less explored.The paper proposes a deep architecture for mobile-phone typing dynamics as complex event sequences.Source: §5 · p. 8
- Prior multi-view approaches include tensor-product fusion with low-rank feature interactions and Factorization Machines for multi-field categorical data.Other work incorporates multiple views into subgraph mining [8] and deep learning.Source: §5 · p. 8
6 CONCLUSION
The paper concludes that mobile-phone metadata could predict mood-disorder presence and reports 90.31% accuracy for DeepMood. Passive data collection may support earlier interventions and deeper understanding of daily activity effects.
6 CONCLUSION
- DeepMood achieves 90.31% prediction accuracy, and passive mobile-phone metadata could support earlier interventions and deeper understanding of mood disturbances in daily activities.The paper states that such interventions may reach more patients earlier in their mood episodes.Source: §6 · p. 8