276

Papers

10/10

Roadblocks Active

0

Connections

This study asks whether brain scans taken before treatment can predict which patients will cognitively improve after VR-based therapy — a key step toward precision prescribing rather than one-size-fits-all rehab programs. Finding that neuroimaging biomarkers do predict benefit means clinicians could eventually route patients to VR therapy only when their brain profile suggests it will work. This matters because cognitive deficits in mood and psychotic disorders are often undertreated, and VR therapy is resource-intensive, so matching the right patient to the right tool has real clinical and economic implications.

██████████ 0.9 digital-therapeutics Peer-reviewed

Read

The study systematically tests whether acoustic speech features — such as shimmer (amplitude irregularity) and jitter (pitch irregularity) — reliably track symptom severity in depression, anxiety, and ADHD across both controlled benchmark datasets and real clinical settings. Crucially, the relationships hold up outside the lab, which is the main obstacle stopping speech-based tools from clinical adoption. If these features are truly stable biomarkers, a microphone could eventually serve as a low-cost, passive symptom monitor requiring no patient self-report.

██████████ 0.9 depression-biomarkers Preprint

ReadSaveConnections

A recurring problem in EEG-based depression detection is that datasets are small, so models are typically improved by generating fake data — a practice that can introduce artefacts. This paper sidesteps that by training an unsupervised model on healthy brainwave data alone, then using how abnormal a patient's EEG looks relative to that healthy baseline as an input signal to guide classification. The added benefit is a spatial adaptation module that handles the messy real-world problem of different EEG headsets having different electrode placements, making multi-site pooling more feasible.

██████████ 0.9 depression-biomarkers Preprint

ReadSaveConnections

Deep learning models can classify depression from EEG with reasonable accuracy, but clinicians and regulators need to understand which brain signals are driving predictions before they can trust the tool. This paper applies five different explanation methods to the same model and checks whether they agree — they largely do, consistently pointing to frontal, temporal, and right-hemisphere regions, which aligns with established neuroscience of depression. Convergence across methods is an important credibility marker: it suggests the model is picking up real biology rather than dataset artefacts.

██████████ 0.9 depression-biomarkers Preprint

ReadSaveConnections

This pilot study combines two brain-monitoring technologies — EEG (electrical activity) and fNIRS (blood-flow proxy) — to classify depressive states without relying on patient self-report, which is notoriously unreliable. The framing around detecting latent depression in people who don't recognize their own symptoms is clinically interesting, particularly for older adults where depression and early dementia can present similarly. The study is very small (11 healthy students) and should be treated as a proof-of-concept only, not a validated tool.

██████████ 0.9 depression-biomarkers Preprint

ReadSaveConnections

Digital therapeutic apps fail largely because patients stop engaging with them, but most recommendation algorithms treat adherence as a fixed characteristic rather than something the app's own behavior influences. This paper models adherence as a dynamic state that evolves in response to what the app recommends and whether the patient followed previous recommendations — a much more realistic picture. The resulting algorithm (UCB-BOLD) comes with theoretical guarantees on how quickly it learns the right strategy, which is a meaningful step toward trustworthy adaptive digital mental health tools.

██████████ 0.9 digital-therapeutics Preprint

ReadSaveConnections

Five large language models were tested on 555 real clinical interviews diagnosed with structured psychiatric tools, and performance was inconsistent: accuracy ranged from near-chance (0.49) to reasonable (0.86) depending on diagnosis and model, with Matthews correlation coefficients — a fairer metric than raw accuracy — staying low across the board. The finding of higher depression-classification accuracy in male versus female participants is a concrete bias signal, not just a theoretical concern. These results suggest LLMs are not yet reliable enough for unsupervised psychiatric screening and need demographic fairness auditing before deployment.

██████████ 0.9 digital-therapeutics Preprint

ReadSaveConnections

Most NLP tools for depression assessment look at a single therapy session in isolation, ignoring that depression fluctuates over time and that earlier sessions contain clinically relevant context. EmoTrack combines LLM-extracted clinical signals with a memory mechanism that compresses prior sessions into a compact representation, achieving a 13.5% improvement in depression severity prediction over the best single-session baseline on a standard dataset. The ability to track trajectories rather than snapshots is important for monitoring treatment response and catching deterioration early.

██████████ 0.9 depression-biomarkers Preprint

ReadSaveConnections

This integrative review maps documented psychological and physical harms of mobile phone addiction in adolescents, a population where mental health deterioration over the past decade closely tracks smartphone adoption. While the specific effect sizes are not visible in the available text, a synthesis of this literature is useful for policymakers and clinicians assessing screen-time guidance. The limitations of integrative reviews — no formal risk-of-bias assessment, no registered protocol — mean the findings should be treated as a narrative map rather than quantitative evidence.

██████████ 0.8 youth-mental-health-crisis Peer-reviewed

Read

This paper argues that the reason two identical people in identical situations have different mental health outcomes is not random noise but a structured, time-varying internal state — a weighting of biological, physiological, and neuropsychological factors active at the moment of a decision. The causal framework proposed suggests this state can be targeted directly through intervention, which would explain why some digital therapeutics work for some people at some times but not others. If the latent-state model holds up, it would provide a formal basis for why adaptive, moment-to-moment interventions should outperform static treatment protocols.

██████████ 0.8 digital-therapeutics Preprint

ReadSaveConnections