Researchers built an AI system that plays diagnostic twenty questions to identify rare diseases faster than static checklists.
The tool, called HPOQuest, targets a real gap: more than 300 million people worldwide live with one of over 7,000 known rare diseases, and diagnosis often stalls because patients initially present with only a handful of symptoms, or phenotypes in clinical terms. HPOQuest starts from that sparse set of observed symptoms and keeps a running, probabilistic ranked list of likely diseases. Rather than ordering every possible test up front, it picks the single most informative follow-up question to ask next. Each confirmed symptom updates the disease ranking, and every answer, positive or negative, refines which question gets asked next.
Tested against four benchmark patient cohorts, HPOQuest lifted the chance of naming the correct disease as the top guess by up to 30 percentage points, and pushed the correct answer into the top five results by up to 45 percentage points, compared to baseline approaches working from the same limited starting information. That is a meaningful jump for a field where doctors often start with just one or two symptoms, and where chasing the wrong disease early can send an entire workup in the wrong direction.
It is a benchmark result, not a clinical trial, and no algorithm replaces a geneticist's judgment, but a system that asks the right next question beats one that simply lists every disease matching the first few symptoms.