A conversational medical assistant for patient interviews

The classic medical interview in the office eats up a large part of the consultation: the doctor asks a series of questions, notes the answers, and decides on the next question based on the previous ones. A client from the medical field wanted to take the load off this process — let the patient conduct a preliminary, structured interview in a mobile app (by voice or tap), and show the doctor already-sorted conclusions together with suggested diagnostic directions.

We built a mobile app available natively on Android (Kotlin) and iOS (Swift), offering two parallel interaction modes: voice — the patient answers by speaking, the app recognizes the speech, extracts the answer and passes it to the decision engine; tap — the patient selects an answer from predefined options, ideal for those who prefer privacy or are in a noisy environment. The backend (Node.js + PHP/Laravel) holds the business logic, the ML for answer analysis, and the API for both clients. Doctors get a separate web panel in Vue with a view of conducted interviews, diagnostic suggestions and the ability to accept or correct them.

The voice mode is based on native speech-recognition libraries (Speech Framework on iOS, SpeechRecognizer on Android), with an optional fallback to a cloud service for longer utterances. The recognized text passes through a normalization layer (removing hesitations, phonetic corrections typical of Polish medical terms) and is mapped to predefined answer categories — so that the decision engine operates on a uniform representation regardless of whether the patient tapped an option or spoke it.

The heart of the app is the ML module, which:

1. Selects the next question based on the answers so far — a decision tree trained on data prepared by the client's medical team, optimizing the reduction of diagnostic uncertainty in as few steps as possible.
2. Classifies signals from the answers into symptom categories (digestive, respiratory, neurological, etc.) and weights them by duration, intensity and the patient's prior conditions.
3. Suggests diagnostic directions — generates a list of the most likely conditions consistent with the collected interview, along with a match score. The result is not a diagnosis — it is input material for the doctor leading the consultation.

The patient conducts a preliminary interview in the app at a pace comfortable for them, and the doctor gets a structured report instead of raw notes. The app is in both stores (App Store, Google Play), and the doctor module is used by the client's medical teams.