Proposed interventions to facilitate patients to come to mammography with higher probability
The interface walks from the data to a concrete action: see the cohort, train a model, score one patient, decide what to do about her.
- Home — the cohort at a glance: how many invitations and patients, what share never respond, and how non-response varies by age band, by registration with a GP or gynaecologist, by how often she sees a doctor at all, and across VZP branch areas
- Risk model — choose the label horizon and the features, train the model, and judge it: precision and recall at a chosen threshold, the ROC curve, and which features it leans on. The trained model is saved and every other page uses it
- Patient risk — enter any patient profile and get her probability of responding, the five features driving that number, and her response speed across the 3, 6 and 12 month horizons
- Interventions — the non-attending patients ranked by risk. Pick one and the page proposes which of her modifiable attributes to change and how much risk that removes, then drafts three reminders for her and an email to each of her doctors
- Refusal profile and Responder profile — portraits of the women who never attend and of those who reliably do, described on the features the model found most important
- Refusal/responder comparison — both groups on a single radar, so the behavioural difference is one picture
- Had before
- nothing
- Use of AI
- - Gradient-boosted decision trees (CatBoost) — risk model predicting whether a woman will be screened after an invitation
- Three parallel models (3 / 6 / 12 months) — the differences between them segment her by response speed: prompt, late (a timed reminder works), or unreachable by reminders at all, where the route has to be her doctor
- Feature importance across the cohort to find which levers exist, and SHAP values per patient to explain one woman's risk
- Counterfactual search over actionable features — scores modified copies of her record and returns the sequence of changes that lowers risk most, restricted to things an intervention can actually move
- OpenAI API — three escalating reminders for her, plus an email to her GP and gynaecologist, generated from her full record with no identifiers sent - Impact
- Data-driven interventions to reduce non-attendance at mammography screening, save lives, and reduce the administrative burden on health insurance companies and general practitioners.
- Who benefits
- Patients and their families, Doctors and nurses, Hospitals and care providers, Health insurers
- InterSystems technologies
- Yes — Implemented the functionality to retrieve real data from a FHIR API server instead of using a plain CSV file. A local FHIR server with test data has also been set up, but it is not currently used because the data was provided in CSV format.
- Team
- mammo_prg — Alex Yudzin, Varvara Shorina, Aleksei Pashchenko
- Challenge
- From invitation to mammogram · Prague