We built an AI system for a global oncology research network that reads clinical notes, matches patients to trial criteria and explains each match. Enrolment became 4x faster, with 65% fewer screen failures.
Industry: HealthcareService: AI & AutomationUpdated
Overview
The challenge
Trial inclusion and exclusion criteria sat in long, complex PDFs. Patient details lived in unstructured clinician notes spread across 14 separate EHR systems. Matching the two meant research staff reading charts by hand.
With more than 50 trials running at once, that manual work slowed enrolment, took most of the research team's time and let unsuitable patients reach screening.
Under-enrolment put trials at risk and delayed the route to market for new therapies.
Research staff spent 70% of their time on administrative data entry instead of patient care.
Inaccurate manual matching caused high screen-failure rates at trial sites, wasting expensive lab resources.
The solution
Medical entity recognition (MER) pulls TNM staging, biomarkers and ECOG scores out of clinical notes inside a secure workflow.
An AI knowledge graph compares each patient profile with trial protocols and flags the most likely matches.
Every match comes with a written justification, so clinicians can see the evidence behind each enrolment recommendation.
Before and after
Before
Human review
Analysts read thousands of clinical notes every month.
After
Automated parsing
Oncology biomarkers are extracted from free-text notes as they arrive.
Before
Policy-based
Staff training was the main guard against patient data leaking during review.
After
Built into the system
Unneeded patient data is removed automatically, and storage is built to HIPAA requirements.
Before
12-14 days
From a patient's diagnosis to a trial recommendation.
After
Under a minute
Match notifications reach clinicians almost at once through secure EHR alerts.
Key features
Investigators are alerted early when a new trial protocol fits a patient already in the system.
Business impact
92% matching accuracy
All processing runs in temporary, memory-only environments, so no patient data is left stored where it could leak later.
Business impact
100% memory-only processing
Clinical coordinators get an alert the moment an EHR update means a patient meets a trial's requirements.
Business impact
4x enrolment speed
Results
4x Matching Speed
Compared with manual chart review, AI matching cut the time from diagnosis to trial recommendation sharply.
Before: ManualAfter: 400% Faster
65% Screen Failures
Automated checks mean only patients who meet 100% of a trial's criteria reach the lab stage.
Before: Error-ProneAfter: 65% Lower
400ms Inference
The language AI system processes thousands of medical notes per second without slowing clinical workstations.
Before: < 2sAfter: 400ms
Faster delivery
These tested components saved about 14 weeks of setup work. We spent that time on trial-specific matching logic and the explanations clinicians read.
Our tested privacy module anonymised patient details before any note reached the matching models.
A reusable collection module pulled patient notes from the 14 EHR systems into one secure pipeline.
An existing monitoring platform tracked the matching pipeline and model response times while clinical teams used it.
Pre-built oncology features gave the models TNM staging, biomarker and ECOG fields to work with from the start.
Trust and control
Every workflow runs inside the client's private cloud, so patient data never leaves its environment.
HIPAA-GATEDEach match includes SHAP explanations and citations back to the original clinical notes.
VERIFIABLE DATACloud automation with SOC 2-ready controls patches vulnerabilities on medical edge devices automatically.
ENTERPRISE SECURITYThe client owns 100% of the oncology knowledge graphs and the fine-tuned language model settings.
100% OWNEDThe system reads unstructured clinical notes, extracts TNM staging, biomarkers and ECOG scores, and compares each patient with trial protocols through a knowledge graph. That replaced manual chart review, which took 12-14 days from diagnosis to a trial recommendation, and made enrolment 4x faster.
Screen failures fell by 65% because automated checks pass only patients who meet 100% of a trial's criteria to the lab stage. Before, inaccurate manual matching sent unsuitable patients to screening and wasted expensive lab resources. Each match now comes with an explanation that clinicians can check.
Patient data stays inside the client's private cloud and is processed in temporary, memory-only environments. Unneeded details are removed automatically before matching, and storage is built to HIPAA requirements. A HIPAA compliance lead was part of the team, and every match cites the source clinical notes for audit.
If trial criteria and patient records sit in separate systems, we can help you build matching that explains every recommendation to clinicians.