MedAssist

1. 01.

   One tone for everyone.

   Symptom checkers speak in a single register. The empathy study found two real dimensions in how people respond to others' feelings, and a spread of people across them.

2. 02.

   Worry changes how people write.

   Anxious people type in short bursts, repeat themselves and shout. A fixed-tone assistant ignores all of it.

3. 03.

   Safety cannot depend on tone.

   Warmth is only safe on top of rules that do not move: no doses, no verdicts, and a hard stop for emergencies.

4. 04.

   The asker is often not the patient.

   Carers describe a child or a parent, and what must never be suggested changes with who it is about.

The tone model stands on an empathy study of 260 people, accepted at HCICC-2026.

• Survey

  The 16-item Toronto Empathy Questionnaire plus age and gender, as an online form.

• Sample

  277 collected, 260 analysed. 57% male, 41% female, 2% other. Mean age 24.6.

• Analysis

  14 methods: reliability, factor analysis, item response theory, invariance, network psychometrics, latent profiles and clustering.

• Regulation

  Telemedicine Practice Guidelines 2020, CDSCO software guidance, the DPDP Act 2023, and Anthropic's policy for health content.

• Pattern study

  A clean-room study of how Ada Health runs a symptom check, from public sources only.

1. 01.

   Empathy has two sides.

   Affective, caring items form one factor (19.3% of the variance). Inhibition items, like steering away from someone's problems, form another and take the total to 31.5%.

2. 02.

   Women scored 4.07 points higher.

   The effect is moderate (r = 0.287). Age was unrelated to score.

3. 03.

   Two profiles, and a small low-reactive group.

   56% are Resonators and 44% Pragmatists. A low-reactive group of 31 people (11.9%) wants control and rational delivery.

4. 04.

   The scale does not work the same for everyone.

   12 of 16 items behaved differently for men and women, so the tone model stays adjustable and gender-neutral by default.

A fourth input is who the chat is about. For a child the app never suggests any medicine, a baby is always severe and a child at least moderate.

• 1. Emergency floor

  What happens

  A fixed phrase list across 13 categories, with misspellings and negation handling. If it trips, the Crisis screen opens and the model is never called.

• 2. Read the person

  What happens

  Five passive signals update the distress score and the Mode.

• 3. Read the symptom

  What happens

  Symptoms are extracted into structured data, then looked up for Severity. Extraction never decides the floor.

• 4. Reply

  What happens

  The model gets Mode, Register, Severity, age band, who it is about, the memory so far and the rulebook.

• 5. Output guard

  What happens

  Every reply is checked for doses, verdicts, brand names and child-medicine rules. A failing reply is replaced by a safe template.

• 6. Render

  What happens

  Short bullets, bold key phrases, a see-a-doctor line on moderate and severe, and a persistent disclaimer.

• Engine unit tests

  What it covers

  Floor, rulebook, output guard, severity, memory, signals, rate limits

  Status

  25 test files

• Acceptance scenarios

  What it covers

  24 scenarios from a first-time user to question limits

  Status

  Written first, run on every change

• Browser walkthrough

  What it covers

  Onboarding, chat and Crisis on a real page

  Status

  Run on Days 4 and 5

• Clinician review

  What it covers

  Safety rules, emergency triggers, 134 tiers, sample replies

  Status

  Pack written, reviewer not yet named

• The research is a survey of 260 people, mostly aged 21 to 25 and mostly students. It describes how they answer a questionnaire, not how they behave when ill.
• The three starting tones come from a scoring key on a cleaned set of 271 responses. They are a labelling heuristic, not the published profile analysis.
• Live replies come from a language model. The rules around it are deterministic but the wording is not, so tests check rules and tiers, not exact sentences.