Turn ADHD progress notes into traceable scores.

Read the note. Keep the evidence.

Tracking progress across ADHD sessions means reading notes written in each clinician’s own style. marc1-Pulse is a small open-weight model trained to turn that free text into a structured reading, with a sentence from the note behind every score.

The model returns eight fields: inattention; hyperactivity and impulsivity; emotional regulation; school or work; home and family; social; medication adherence; and side effects. Each is either null, when the note does not document that area for the current period, or a score from 0 to 3 with a verbatim supporting quote. It also lists safety concerns and informants.

Deterministic code computes the downstream measures: a CGI-Severity proxy, change from baseline, responder status, reliable change, and the trend across sessions. The model extracts and quotes; the code performs the arithmetic.

For example, the study’s illustrative note includes “Adherent to dexmethylphenidate ER, taken each AM with breakfast.” The example output assigns medication adherence a score of 0 and preserves that exact sentence as evidence. Areas the note does not document remain null. The example is explicitly illustrative and does not describe a real person.

Evaluate on public clinical writing.

The study reports 1,518 openly licensed clinical documents added to training, from MTSamples transcriptions and PubMed Central open-access case reports. Every retained example used the production schema and source-verified quotes. Training covered eight areas, the full severity range, and seven documentation styles, including intake, discharge, collateral contact, and routine notes.

At most 120 documents with nothing to score were kept from each source. This taught the model that a note can be empty without making empty notes dominate the training. No training document shared a source article with a development or test note.

The test reference was a three-model panel: Claude Opus 5.5, Claude Sonnet 5.5, and Qwen3.8-27B. Models outside the panel were scored only on cells where all three agreed. Each panel member was scored against cells where the other two agreed. The panel disagreed on 11.6% of cells; those contested cells were not scored.

Development results used Opus alone as the main reference and are indicative. Most evaluation documents were journal case reports rather than a clinic’s routine session notes, an important distinction when interpreting the results.

Area score is an agreement metric, not an accuracy percentage.

The area score combines three checks: whether the model scored the documented areas, how close each 0-3 score was to the reference, and whether a supporting quote exists in the note.

A score with no verbatim quote is treated as null before grading. A cell enters the calculation whenever either the model or the reference scored it. If just one side is null, the cell earns zero. If both scored the area, it earns partial credit by distance:

Cells where both sides are null are excluded. The references scored about 1.8 of eight areas per note, so counting jointly empty cells would reward a model that reports nothing. Under this definition, that model scores zero.

Safety recall is measured separately. It is not folded into the area score, because missing a safety concern is a different kind of error.

Reported performance on 178 held-out notes.

The reported paired difference against Sonnet was -0.002, with a 95% interval from -0.043 to +0.039. Its lower bound cleared the pre-specified non-inferiority margin of -0.05 on this test. These are the document’s reported paired statistics; the displayed model scores are rounded.

Pulse scored 0.903 on notes whose source article had a psychiatric title and 0.861 on the 2025-2026 case reports. The study reports serving costs of $1.00 per 1,000 notes for Pulse, $1.50 for Haiku, $11.70 for Sonnet, and $25.87 for Opus. Those are the study’s cost comparisons, not current service quotes.

Two clinicians separately scored 60 test notes. Their agreement with each other was 0.84; Pulse’s agreement with them was 0.82 and Sonnet’s was 0.83. This provides a limited human-reference check alongside the model panel, rather than establishing performance across a clinical population.

A flag means review, not an automated decision.

Pulse extracts concerns involving self-harm, harm to others, abuse, and medication misuse. A deterministic keyword net runs alongside the model and can only add flags. Either source can send a note to review.

The keyword net improves coverage but does not eliminate misses. The test still left safety concerns unflagged, and the pipeline also produced false alarms. The paper’s comparison applied the same keyword net to all models; its chart reports 94.1% recall for Sonnet and 97.5% for Haiku with that net.

A source quote that cannot be found, or a reading too thin to support a change from baseline, also triggers review. The intended workflow keeps these exceptions visible to the clinician instead of silently presenting a complete-looking score.

Real writing changed the result more than model size alone.

An early model looked strong on notes from its own training distribution but left 62% of areas scored by the reference empty on the development notes. It recognized explicit, full-sentence descriptions and missed shorthand or passing mentions.

A larger 8B model did not solve that gap. Changing the base to Gemma 4 E4B reduced the omitted-area rate to about 40%. Adding 630 real clinical documents reduced it to 23.8%. The final four-adapter reading reduced it to 10.3% on development notes.

The final system reads each note with four LoRA adapters on the same base model, two using a decoding penalty against empty answers. It scores an area when any adapter scores it and flags a concern when three of four do. The adapters miss different areas, so their combined reading recovers coverage.

That improvement comes with a tradeoff. Against Opus’s development reading, Pulse invented a score 8.0% of the time versus 6.8% for Sonnet; on cells where the whole panel agreed, the reported rates were 1.0% and 1.3% respectively.

The source also records a curation failure: allowing one label disagreement per note taught the model to ignore brief mentions. Correcting 1,228 disputed cells reversed that regression. Reinforcement learning on one candidate increased false alarms beyond the pre-set limit, and no individual development intervention was enough on its own.

Keep the reference and the clinical boundary visible.

Much of the reported evaluation is agreement with other models. The study checked 42 selected, blinded disagreements and found that neither competing reading was always correct. Those selected cases expose reference weaknesses; they do not estimate how often a reference is wrong.

The team tried 17 development variants. The retained variant still failed the -0.05 non-inferiority threshold on development data under both the single-model and panel references. The positive non-inferiority result belongs to the separate 178-note held-out test, not every dataset examined.

The study does not demonstrate diagnostic validity, treatment effectiveness, or prospective clinical outcomes. Public case-report text differs from routine documentation, and the human-reference check covered 60 notes. A new organization’s note styles and review workflow remain a separate evaluation setting.

Owning the model weights can keep the data path and model version under the organization’s control. That does not remove the need to validate the extraction task, monitor documentation changes, and retain clinician review.

Full study and data notes.

The ten-page October 2026 case study contains the full experimental sequence, confidence intervals, safety breakdowns, illustrative output, and reference checks. Model names and comparison settings here follow that document.

Training sources identified in the study are MTSamples (CC0) and PMC open-access case reports (CC BY), with attribution in the dataset card. Evaluation case summaries come from PMC-Patients (CC BY-NC-SA). The reference models are Claude Opus 5.5, Claude Sonnet 5.5, and Qwen3.8-27B; Haiku 5.5 is a comparison baseline.

The downloadable PDF is the original source, unchanged. This page reports its study results; it does not publish patient records or the underlying evaluation dataset.