ADOS Clinical Dashboard (Thesis)

ADOS Clinical Dashboard (Thesis)The final deliverable of my Bachelor's thesis on autism-screening research: an offline clinical dashboard for reviewing a toddler's ADOS-2 session on one synchronized timeline.This dashboard runs entirely locally, with no outbound network calls: it's the presentation-facing layer of a much larger thesis pipeline built for GDPR Art. 9-sensitive health data of minors. The IMU signal processing, feature engineering, the sensor-only ADOS (Autism Diagnostic Observation Schedule) risk model (a RandomForest trained for the thesis's data-science analysis, evaluated out-of-fold) and the statistical analysis behind every number shown here all run upstream on the same machine. Clinician notes are summarized by a local, extractive text model that lifts sentences verbatim, never a cloud LLM, chosen after generative summaries were found to fabricate clinical claims. The thesis itself has already been completed and delivered; this dashboard is the part still being refined and polished.

Synchronized Patient Detail: video playback linked bidirectionally to an interactive label timeline and the raw IMU signal plot; toy and camera angle are selected independently per session.

Sensor-Only Analysis (new): runs the sensor-only ADOS risk model on a child's motion data alone, benchmarked against the cohort and same-risk peers, with three overlaid movement timelines comparing automated detection to the human record.

Data-quality transparency (new): a saturated sensor, broken recording or low sample coverage surfaces an explicit warning next to the prediction, instead of a silently wrong number.

Clinician Notes with extractive AI summary (new): a rich-text editor (multiple notes/clinicians per patient, exportable to docx/pdf/txt) that summarizes notes by lifting sentences verbatim, after generative summaries were found to fabricate clinical claims.

Toy Analytics: cross-patient view of action counts, intentionality (child-initiated / accidental / educator-led) and action diversity per toy.

Cohort data-quality overview: surfaces per-child completeness directly (missing labels, unreadable data, sensor issues) rather than hiding it behind an aggregate.

Tech stack: Python, Dash/Plotly, Flask, Pandas, scikit-learn, Quill.jsA project by Manuel Cattoni.