AI Clinical Trial Search Engine for Patient Matching
~75%
2–4 months
Not disclosed
Patients couldn't find relevant clinical trials without physician involvement. Legacy search relied on outdated text matching that missed medical synonyms, and the system had to rank 65,000+ active trials within sub-5-second response times.
The team had originally tried to maintain a condition mapping by hand. That approach covered less than 5% of their target domains, and doing it properly would have required an estimated 5-10 full-time people.
PressW built a custom NLP engine with automated medical-condition hierarchy mapping, a multi-level search architecture (instant results with background deep matching), and AI models fine-tuned to clinical-trial eligibility criteria.
The team automated condition-relationship mapping (no manual taxonomy), pre-calculated key condition families for common searches, and implemented tiered search levels delivering immediate results while deeper analysis ran concurrently.
Clinical-trial matching platforms and patient-facing healthcare search applications that need medical-synonym awareness and sub-second relevance at scale.
- Registry of 65,000+ active clinical trials with eligibility criteria (source data)
- Medical condition dataset spanning more than 1M conditions
- Pre-calculated condition families for common searches
- Patient-facing search application
- Trial registry ingest to automated condition-relationship mapping to indexed hierarchy
- Patient query to tiered search: immediate results returned while deep analysis runs concurrently
- Pre-calculated condition families short-circuiting common searches
- Fine-tuned eligibility models scoring patient-to-trial relevance
1,000+ Manual Hours Saved
65,000+ Active Trials Indexed
Matching in Under 5 Seconds
The hardest part operationally was building test data to validate whether answers were good enough in a clinical setting, since that data is difficult to obtain. The hardest part technically was the level of concurrency required, and then cross-checking those concurrently derived results to decide which combination of answers was best.
Bryson Greenwood
ClinicalNet replaced outdated keyword matching with a custom NLP engine. Condition relationships were mapped automatically rather than through a hand-built taxonomy, common condition families were pre-calculated, and a tiered search architecture returns immediate results while deeper analysis runs concurrently. Matching now completes in under five seconds across 65,000+ active trials.
Custom NLP models fine-tuned to clinical-trial eligibility criteria, combined with custom-trained NER models, embedding models, and automated medical-condition hierarchy mapping across a multi-level search architecture. No off-the-shelf model is named.
Automated matching removed 1,000+ hours of manual trial-search work, the engine indexes and ranks 65,000+ active trials across more than a million conditions, and patients get relevant matches in under five seconds with core results in under one.
PressW reports meaningful results within weeks. The work covered automated condition-relationship mapping, pre-calculating key condition families, and implementing tiered search levels. The team had first tried a hand-built condition mapping, which covered under 5% of target domains and would have needed an estimated 5-10 full-time people to maintain properly.
Clinical-trial matching platforms and patient-facing healthcare search applications that need medical-synonym awareness and sub-second relevance at scale.
Have a similar challenge?
