The Challenge: Thousands of Patients, Hundreds of Eligibility Criteria
A specialist cancer research centre affiliated with a London teaching hospital runs 200 active clinical trials simultaneously. Each trial has between 15 and 60 eligibility criteria spanning diagnosis type, staging, prior treatment history, biomarker status, organ function thresholds, and demographic requirements. Every month, 3,000 newly referred oncology patients arrive, and the research nurse team must determine which patients qualify for which trials. Currently, a research nurse spends an average of 40 minutes per patient manually cross-referencing medical records against trial protocols. At 3,000 patients monthly, the maths does not work — the team can only screen a fraction of referrals, and eligible patients miss trials they could have benefited from.
Automating this matching requires processing unstructured clinical documents — referral letters, pathology reports, imaging summaries, blood results — and comparing extracted patient attributes against structured eligibility criteria. The data involved is profoundly sensitive, and the centre’s R&D governance board will not approve processing that routes patient records through non-GDPR-compliant infrastructure.
AI Solution: Document AI + LLM for Eligibility Screening
The matching pipeline combines two AI capabilities. First, document AI extracts structured information from clinical records: diagnosis codes, staging (e.g., T3N1M0), ECOG performance status, lab values (creatinine, neutrophil count, liver enzymes), and prior treatment regimens. PaddleOCR handles scanned documents, while NLP models extract entities from digital text.
Second, an open-source LLM receives the extracted patient profile and each trial’s eligibility criteria, then determines match/no-match with explanatory reasoning. This is not simple keyword matching — criteria often involve nuanced clinical logic like “no prior immunotherapy within 6 months unless discontinued due to toxicity” that only a language model can parse reliably.
GPU Requirements: Batch Processing Thousands of Matches
Each patient-trial match evaluation involves feeding the LLM a structured prompt containing the patient profile (300-500 tokens) and trial criteria (200-800 tokens), then generating a match assessment (100-200 tokens). With 3,000 patients and 200 trials, the theoretical maximum is 600,000 evaluations — though pre-filtering by cancer type typically reduces this to 30,000-50,000 relevant patient-trial pairs per month.
| GPU Model | VRAM | Evaluations per Hour (Mistral 7B) | 50,000 Pairs |
|---|---|---|---|
| NVIDIA RTX 5090 | 24 GB | ~800 | ~63 hours |
| NVIDIA RTX 6000 Pro | 48 GB | ~1,100 | ~46 hours |
| NVIDIA RTX 6000 Pro | 48 GB | ~1,300 | ~39 hours |
| NVIDIA RTX 6000 Pro 96 GB | 80 GB | ~2,000 | ~25 hours |
An RTX 6000 Pro on GigaGPU dedicated hosting processes the monthly screening load in approximately one day, delivering match reports to research nurses by the second working day of each month. Daily incremental processing of new referrals completes in under an hour.
Recommended Stack
- PaddleOCR for digitising scanned pathology reports and referral letters into machine-readable text.
- spaCy + medCAT for clinical named entity recognition — extracting diagnoses, medications, lab values, and procedures from unstructured text.
- Mistral 7B-Instruct or LLaMA 3 8B for eligibility evaluation, served via vLLM for batched throughput.
- ClinicalTrials.gov API integration for automatic ingestion of updated eligibility criteria as trials open and amend protocols.
- Streamlit dashboard for research nurses to review AI-generated match reports, override decisions, and track screening metrics.
An AI chatbot interface allows clinical research associates to ask natural language questions: “Which of our breast cancer patients with BRCA1 mutations are eligible for open Phase II trials?” — with the LLM querying both the patient database and trial registry in real time.
Cost vs. Alternatives
Commercial trial matching platforms from vendors like Deep 6 AI or TrialJectory charge annual licensing fees of £100,000-£300,000 and require data to flow through their cloud infrastructure. Self-hosting on dedicated GPU provides equivalent matching capability at a fraction of the cost while keeping every patient record on identifiable UK infrastructure.
The clinical impact dwarfs the cost discussion. Research centres using AI-assisted matching consistently report 30-50% increases in trial enrolment rates — meaning more patients access experimental treatments and trials reach statistical power faster, accelerating the path to new therapies.
Getting Started
Select five high-enrolment trials and retrospectively test the AI matcher against the last six months of manual screening decisions. Measure sensitivity (did the AI catch every patient the nurses identified?) and specificity (did it avoid false matches that would waste screening appointments?). Most teams achieve 90%+ sensitivity within the first tuning iteration.
GigaGPU delivers private AI hosting with the compute power clinical trial matching demands and the GDPR compliance research governance boards require. Process thousands of patient-trial pairs without a single record leaving UK data centres.
GigaGPU’s dedicated servers process eligibility screening at scale while keeping patient data sovereign. No per-match fees, no shared tenancy.
View Dedicated GPU Plans
