Building an AI Matching Algorithm with 50+ Decision Criteria

The Problem

Healthcare provider matching is one of the most complex recommendation problems in technology. Unlike e-commerce recommendations (where a wrong suggestion costs nothing) or content recommendations (where a bad video just wastes a few minutes), healthcare matching has real consequences. A patient matched with the wrong therapist may not get the treatment they need, potentially worsening their condition. A patient matched with a provider who does not accept their insurance faces surprise bills. A patient matched with a provider too far away may skip follow-up appointments, undermining treatment outcomes. The dimensionality of healthcare matching far exceeds typical recommendation systems. A patient searching for a therapist must consider: specialty alignment (CBT vs DBT vs psychodynamic), condition expertise (anxiety vs depression vs trauma), insurance acceptance (hundreds of plan variations), location and commute tolerance, availability that matches their schedule, language preferences, gender preferences (many patients prefer same-gender therapists), telehealth vs in-person preference, pricing for cash-pay patients, years of experience, patient review scores, cultural competency, and dozens of other factors. Existing platforms handle this complexity poorly — Zocdoc primarily filters by insurance and location, ignoring most of the factors that actually determine a successful patient-provider match. The result is a frustrating experience where patients must manually evaluate providers against their personal criteria, leading to the 26-day average booking time.

Results

The 50-criteria matching algorithm achieved a 94% first-match satisfaction rate — meaning 94% of patients who booked through OpenMyPro's top recommendation reported being satisfied with their provider choice. This compares to approximately 60% satisfaction with provider choices made through traditional search-and-compare methods on competing platforms. The algorithm's 0.91 AUC (area under the ROC curve) indicates strong predictive accuracy for booking completion, and the model continues to improve as more interaction data is collected. The rebooking rate of 67% — the percentage of patients who return to OpenMyPro to book another appointment (either with the same or different provider) — validates that the matching quality creates lasting platform loyalty. The average match computation time of under 2 seconds enables the 33-second total booking experience, with the algorithm running as a serverless function that scales automatically during peak hours (Monday mornings and evenings are highest volume). The technical sophistication of the matching algorithm also became a key differentiator in investor conversations and partnership discussions with health systems. Several hospital networks expressed interest in licensing the matching technology for their internal referral systems, opening a potential enterprise revenue stream that could complement the consumer marketplace.

Related Content