Deep Learning Model Flags Early Signs of Problematic Digital Habits in Older Adults

Excessive smartphone and internet use is usually framed as a problem of the young, but a new study argues that older adults may be quietly developing risky digital habits that go unnoticed by conventional screening. Researchers have built a deep learning framework, called DL-MFE-EW, that mines everyday device usage logs to predict which older users are drifting toward problematic digital behavior — and it issues its warning roughly a week before high-risk patterns set in. On a real-world screen-time dataset filtered to users aged 60 and above, the model reached 0.947 prediction accuracy and 0.93 early-detection sensitivity, outperforming a series of machine learning baselines.

The work, published in Discover Artificial Intelligence, addresses a gap that the authors say has been neglected. Most computational research on internet addiction has concentrated on younger populations, and the assessment tools traditionally used with seniors — self-report questionnaires and clinician-administered screenings — suffer from memory bias, poor time estimation and a failure to catch early warning signs. Older adults also present a harder detection problem for structural reasons: variable daily routines, age-related cognitive changes, shifting social networks and health complications all reshape how digital habits look in this group, so models tuned on younger users do not transfer cleanly.

At the heart of DL-MFE-EW is a hybrid neural architecture that combines convolutional neural networks (CNNs), long short-term memory networks (LSTMs) and an attention mechanism. The pipeline begins with a data acquisition layer that gathers behavioral streams such as daily screen time, application interaction counts, device unlock frequency and notification activity. Because these streams arrive at different sampling rates, the system aligns them onto a common timeline using interpolation-based resampling, fills gaps with temporal autocorrelation-aware imputation, and normalizes every feature with z-scores so that no single modality dominates learning simply because of its units or magnitude.

The CNN component then operates on the resulting behavioral matrix, sliding multi-scale convolutional filters of 3×3, 5×5 and 7×7 size across the data to detect local cross-modal patterns — for example, bursts of device unlocks coinciding with rapid app switching or abrupt changes in psychometric-style indicators. Pooling layers compress these activations while preserving the behavioral signals that matter. The LSTM stream takes over from there: fed with the CNN’s feature embeddings in temporal order, its gated memory cells track escalation dynamics, periodicity and long-range dependencies, such as a slow multi-day rise in evening screen use or cyclical sleep-disrupting sessions. A bidirectional variant processes the sequence in both directions, letting the model relate earlier usage to later outcomes and vice versa.

The attention layer is what turns these representations into an interpretable early warning. Rather than averaging all temporal states equally, the model computes alignment scores for each time step, normalizes them into a probability distribution, and builds a context vector weighted toward the moments most relevant to risk — late-night compulsive use, for instance, or sessions paired with negative emotional indicators. The fused representation feeds a classifier that outputs a continuous risk score, a categorical alert level and an uncertainty estimate. An early-warning index is computed as the attention-weighted sum of behavioral deviations over the observation window, and attention heatmaps paired with SHAP attributions are meant to translate model alarms into actionable signals for caregivers, such as more wake-up calls or reduced overnight screen exposure.

Because genuinely problematic usage is rare compared with normal patterns, the training procedure had to guard against the model simply defaulting to the majority class. The researchers applied class-weighted loss so that errors on the rare risk category counted more heavily, and they evaluated performance with per-class precision, recall and F1-score alongside confusion matrices. Ground-truth labels were generated by rule-based annotation rather than clinical diagnosis: sessions were tagged as normal use, moderate risk or possible digital addiction when they crossed thresholds drawn from the behavioral health literature, such as six or more hours of daily screen time or more than 80 device unlocks per day. Early-warning labels were assigned one week before predicted high-risk states, and a temporal separation between input data and labels was imposed to prevent the model from simply memorizing the signals used for annotation.

The experimental dataset was the publicly available Kaggle Screen Time and App Usage Dataset (iOS/Android), from which only users aged 60 or older were retained. Data were split at the user level — 70 percent training, 10 percent validation, 20 percent testing in the primary protocol — so that no individual’s sessions appeared in more than one partition, a subject-independent evaluation designed to reflect real-world generalization to unseen users. Experiments were replicated five times, and hyperparameters such as learning rate, batch size, dropout and the number of attention heads were tuned by grid search on the validation set. The final architecture used 1-D convolutional blocks, a bidirectional LSTM with 128 hidden units, and a four-head additive attention layer, optimized with Adam and binary cross-entropy loss.

Against baselines including ML-EEG, S-Feat-ML, APPLog-ML and Poll-Reg, DL-MFE-EW led on nearly every metric: 0.947 accuracy versus 0.884 for the best baseline, 0.936 precision, 0.923 sensitivity, an F1-score of 0.93, an AUC-ROC of 0.90 and an early-warning sensitivity of 0.93 (reported as 0.93 plus or minus 0.02 with bootstrap confidence intervals and Wilcoxon signed-rank significance testing). The model also achieved a behavioral drift detection rate of 91.5 percent, meaning it caught gradual shifts in usage patterns — creeping screen-time growth, longer app sessions — that sudden-change detectors missed. Ablation tests underscored the design choices: removing the attention layer dropped accuracy to 89.2 percent, and removing the CNN feature extractor cut the F1-score to 0.845. Calibration analysis yielded a Brier score of 0.08, and decision curve analysis suggested the greatest clinical net benefit at intervention thresholds between 0.6 and 0.8.

Several caveats deserve emphasis. The ground-truth labels are behavioral proxies, not clinical diagnoses of addiction; the authors explicitly state that a formal diagnosis would require a trained psychologist and standard diagnostic criteria outside the scope of this study. Moreover, while the full framework is designed to fuse psychometric assessments (such as mood, anxiety and loneliness instruments) with device logs and circadian proxies, the quantitative results reported here rest solely on the device-usage data — the psychometric and explainability modalities were conceptual components planned for future multimodal deployment. The findings also apply only to the older-adult cohort studied; the authors caution against generalizing to younger users, whose behavioral dynamics differ markedly. Ethical questions around monitoring seniors’ digital behavior, including informed consent and privacy-safe data handling, remain open, and the authors flag limited senior-specific datasets, uneven digital literacy and the difficulty of long-term data collection as practical barriers.

Even with those limitations, the study points toward a plausible future for gerontechnology: lightweight monitoring modules embedded in phones or wearables that continuously score digital habits and alert caregivers before unhealthy patterns harden into dependence. The authors describe near-real-time inference with moderate computational demands, and they outline next steps that include senior-specific behavioral datasets, privacy-aware learning techniques and integration with mobile and Internet of Things platforms. If those pieces mature, the line between passive digital surveillance and protective digital health support for older adults will depend less on the algorithms than on how consent, transparency and clinical oversight are built around them.

Subject of Research: Deep learning-based early detection of problematic digital usage behavior in older adults

Article Title: Deep learning-driven multidimensional feature extraction and early warning method for digital addiction behavior in the elderly

Article References: Zheng, J., & Wang, J. (2026). Deep learning-driven multidimensional feature extraction and early warning method for digital addiction behavior in the elderly. Discover Artificial Intelligence, 6(1), Article 1001. https://doi.org/10.1007/s44163-026-01937-2

Image Credits: AI Generated

DOI: 10.1007/s44163-026-01937-2

Keywords: digital addiction, older adults, deep learning, CNN, LSTM, attention mechanism, early warning system, screen time, behavioral monitoring, digital health, machine learning, gerontechnology