0) Core idea (what we’re replicating—and improving)

• Task engine (CPT family): Go/No-Go + Stop-Signal + Sustained Attention (CPT), with visual + optional auditory streams.

• Activity capture: Webcam-based micromovement tracking (pose/landmark kinetics) ± optional IMU (phone in pocket) to approximate “QbActivity”.

• Outcome metrics: Standard CPT metrics (omission/commission errors, reaction time (RT), RT variability), lapses, post-error slowing, and movement indices (fidget index, head sway, body displacement, micro-saccade proxy).

• Normative scoring: Age/sex-adjusted Q-scores and percentiles using normative modeling (see §5).

• AI layer: Probabilistic severity estimates + explanations (feature importances, counterfactuals), not a standalone diagnosis.

1) Stack & licenses (fully open)

• Front-end task runtime: jsPsych or lab.js (browser-based, low-latency; works on clinic PCs).

• Motion capture: MediaPipe Face/Hands/Pose (WebAssembly via browser) or OpenCV + MediaPipe (Python desktop).

• Back-end & analytics: FastAPI (Python), pydantic, uvicorn.

• Stats/ML: scikit-learn, statsmodels, PyMC (Bayesian), xgboost/lightgbm, shap (explainability).

• Visualization & reports: plotly, weasyprint/reportlab → PDF.

• Packaging: Docker for one-click local installs.

• License: AGPL-3.0 (or Apache-2.0 if you prefer more permissive).

2) Task battery (replicable + extensible)

A. Sustained Attention CPT (15–20 min, adjustable)

• Rare target (e.g., press on “X” with 1:3 to 1:4 non-target ratio).

• Outputs: omissions, commissions, d′ (signal detection), RT mean/SD/CV, lapses (RT > 90th percentile), drift (time-on-task decline).

B. Go/No-Go (5–7 min)

• Outputs: inhibitory failure rate, prepotent bias, post-error slowing, ex-Gaussian RT parameters (μ, σ, τ) for variability.

C. Stop-Signal Task (optional, 8–10 min)

• Outputs: SSRT (stop-signal reaction time), inhibition function (by SSD).

D. Distractor CPT (MOXO-style, optional)

• Background audiovisual distractors (classroom noise, moving objects) with culturally neutral assets.

• Outputs: performance deltas with/without distractors.

All tasks log millisecond timestamps, stimulus IDs, and raw keypress/mouse events.

3) Activity/movement capture (your “QbActivity” analogue)

Webcam-only baseline (no wearables):

• Pose landmarks: head, shoulders, wrists; compute per-frame displacement velocity, jerk, and spectral power of micro-movements.

• Fidget index: weighted sum of (head yaw/pitch variance, wrist trajectory entropy, trunk sway).

• Stationarity drift: long-window variance vs. short-window variance ratio.

• Motion-noise guardrails: face occlusion detection, confidence thresholds; flag invalid segments.

Optional IMU add-on (smartphone in pocket)

• Lightweight PWA reads accelerometer/gyroscope; syncs via WebRTC timestamping.

• Derive RMS acceleration, jerk, and burst frequency—correlate with webcam metrics.

4) Feature set (for both human interpretation and AI)

Attention/Impulsivity (CPT family)

• Omissions, commissions, RT mean/SD/CV, ex-Gaussian τ, sequential effects (AR(1) in RT), post-error slowing, vigilance slope, d′, β.

Activity

• Fidget index, head/torso sway (px/s), movement bursts per minute, micro-pause frequency, entropy of wrist paths, spectral centroid of movement.

Ecological robustness

• Performance with vs without distractors; delta scores summarize susceptibility.

5) Scoring: percentiles & Q-scores via normative modeling

• Normative dataset schema: age, sex, handedness, task version, device, lighting, education band.

• Model: Hierarchical Bayesian Normative Modeling (HBM) or Gaussian Process Normative Modeling:

  • Fits expected value and variance of each feature as a function of covariates (age/sex/device).

  • Individual deviation (z-like Q-score): Q=x−μ^(age,sex,…)σ^(age,sex,…)Q = \frac{x – \hat{\mu}(age,sex,…)}{\hat{\sigma}(age,sex,…)}Q=σ^(age,sex,…)x−μ^​(age,sex,…)​

  • Multi-feature fusion: robust Mahalanobis distance → domain Q-scores (Activity, Inattention, Impulsivity) + Total Q.

• Interpretive bands (proposed, modifiable after validation):

  • 0.0–1.0: within typical limits

  • 1.1–1.4: mildly atypical

  • ≥1.5: atypical / clinically concerning

• Change metrics: ΔQ ≥ 0.5 (half SD) ⇒ clinically meaningful improvement flag.

6) AI-assisted interpretation (decision support, not diagnosis)

• Models: Gradient boosted trees or calibrated logistic regression with probability of clinically significant deviation in each domain.

• Calibration: Platt/Isotonic per domain; reliability curves surfaced in UI.

• Explainability:

  • Global: feature importance bars, partial dependence.

  • Local: SHAP values for the patient; counterfactuals (“If RT variability dropped by 20%, Total Q would reduce by 0.35”).

• Narrative engine: Template + LLM fill for patient-friendly and clinician-grade summaries, gated by strict guardrails (facts only, cite metrics, surface uncertainty).

7) Reporting (clear, audit-safe)

• One-page patient summary: percentile dials for three domains, green/amber/red bands, plain-language explanation.

• Clinician report (PDF): task plots (RT time series, error raster), movement spectrograms, domain Q-scores, ΔQ vs prior visits, QC flags (lighting, face lost %, dropped frames).

• CSV/Parquet export: raw + derived features for research.

8) Data, privacy, consent

• Local-first processing: default on-device; optional encrypted sync to clinic server.

• De-ID: face embeddings discarded; store only numerical landmarks/velocities; SHA-256 of video chunks (no video retention by default).

• Consent tiers: (a) care only; (b) anonymized norms; (c) research sharing (with IRB).

• Audit logs: every model version & parameter set stamped into each report.

9) Validation plan (publishable, pragmatic)

1. Feasibility & reliability (n≈60): test–retest ICC for key features and domain Q-scores over 1–2 weeks.

2. Construct validity (n≈150): correlations with CPT-3/TOVA/IVA-2 and rating scales (ADHD-RS-5, Conners 3).

3. Known-groups validity (n≈200): ADHD vs. non-ADHD; compute AUC per domain.

4. Sensitivity to change (n≈80): pre/post stimulant titration; ΔQ distribution; anchor with clinician CGI-I.

5. Device variance study: cheap webcam vs. HD webcam; adjust normative model if needed.

Power calcs & pre-registration recommended; share scripts to keep it open and credible.

10) Regulatory stance (clear and safe)

• Position as clinical decision support / research tool, not a diagnostic.

• Add Intended Use statement, risk controls, and QC gates.

• If later seeking medical device classification, align code/process with IEC 62304, ISO 14971, and ISO 13485 processes.

11) MVP roadmap (12 weeks)

Weeks 1–2:

• jsPsych CPT + Go/No-Go; FastAPI backend; JSON event logs; webcam capture + MediaPipe pose; minimal UI.

Weeks 3–4:

• Feature extraction pipeline; initial normative bootstrap (healthy volunteers across age bands); simple z/Q-scores; PDF report v1.

Weeks 5–6:

• Add fidget index; distractor module; IMU optional; QC metrics (lighting/face loss).

Weeks 7–8:

• AI v1 (calibrated LR / XGBoost); SHAP explanations; clinician/patient report split.

Weeks 9–10:

• Test–retest study; bug bash; Dockerized installer; offline mode.

Weeks 11–12:

• Documentation, unit tests, telemetry (opt-in), preprint of methods, invite collaborators.

12) Data schemas (practical)

/data/raw/

• events_{uuid}.parquet → trial_id, stim_type, is_target, t_on, t_resp, correct, rt_ms

• pose_{uuid}.parquet → t, landmark_id, x, y, conf

• imu_{uuid}.parquet → t, ax, ay, az, gx, gy, gz

/data/derived/

• features_{uuid}.parquet → flat feature vector

• qc_{uuid}.json → dropped_frames, face_lost_pct, light_score

/reports/

• report_{uuid}.pdf, report_{uuid}.json (for EHR ingestion)

13) Example feature definitions (succinct)

• RT variability (CV): std(RT)/mean(RT)

• Lapses: % trials with RT > mean+2*SD

• Post-error slowing: mean(RT post-error) - mean(RT post-correct)

• Fidget index: w1*var(head_yaw)+w2*var(head_pitch)+w3*path_entropy(wrists)+w4*jerk_rms (weights from training)

14) AI interpretation snippet (logic you can ship)

• Domain probabilities: $P(\text{atypical domain})=\text{calibrated\_model}(features)$

• Total Q: robust combination of domain Q-scores (e.g., trimmed mean).

• Narrative rules:

  • If Q_inattention ≥ 1.5 and RT_CV high → “Marked variability consistent with inattention.”

  • If Q_activity ≥ 1.5 with high fidget index and frequent bursts → “Elevated motor restlessness.”

  • Always surface confidence and QC notes.

15) Ethics & equity

• Bias checks: stratify performance by age, sex, language, device quality.

• Transparent thresholds: publish how bands map to Q/percentiles.

• No black-box edicts: every AI claim paired with visible metrics.

16) Where this goes next (your clinic’s edge)

• Clinic-ready: run assessments before and after stimulant trials, neurofeedback blocks, or school term changes; track ΔQ.

• Research: rapid piloting for ADHD subtyping, digital phenotyping, and treatment personalization.

• Education: patient-friendly visuals demystify “brain fog” vs. inattention, reduce stigma, and anchor shared decisions.

About the Author

✦ Dr. Srinivas Rajkumar T, MD (AIIMS, New Delhi)
Assistant Professor of Psychiatry, Sree Balaji Medical College & Hospital, Chennai
Consultant Psychiatrist, Mind and Memory Clinic, Apollo Clinic, Velachery, Chennai (Opp. Phoenix Mall)

My expertise spans ADHD, neurodevelopmental disorders, and neuromodulation therapies (rTMS, tDCS, neurofeedback, and digital brain-based tools). I am also passionate about integrating AI and open-source methods into clinical psychiatry to enhance diagnostic objectivity and patient outcomes.

📧 srinivasaiims@gmail.com
📍 Chennai, India

Call for Collaboration

If this blueprint nudges your curiosity, don’t wait for perfect—pick a module, build a scrappy v0, and ship. I’m actively collaborating with clinicians, engineers, and researchers on open, auditable ADHD assessment tools.

Email me at srinivasaiims@gmail.com with the subject “Open ADHD Tool — Collab” and a 3–5 line summary of what you’re tackling. When you make progress—code, dataset, validation, or even a negative result—send an update. I’ll credit contributors, review useful PRs, and help test promising ideas in clinic.