RESEARCH
One published paper, four in submission, and an externally evaluated model. Every claim on this site traces to a documented evaluation with its dataset and label type stated. We don't do press releases — we do papers.
Introduces a systematic framework for validating silver labels used in machine learning training pipelines. Demonstrates that instrument validation techniques from psychometrics can detect label noise and quantify measurement error in automatically generated training data.
Examines how different explainability methods (SHAP, LIME, attention weights, integrated gradients) produce divergent explanations for the same clinical predictions. Argues that multi-method agreement scoring provides more trustworthy explanations than any single method.
Proposes methods for validating silver-label quality in clinical trial matching when gold-standard expert labels are scarce. Applies factor analysis and internal consistency metrics to 87.8 million patient-trial pairs.
Analyzes failure modes in neural clinical trial matching systems. Finds that boolean eligibility criteria (age ranges, lab value thresholds, binary diagnoses) account for the majority of ranking errors, and proposes hybrid architectures that combine neural ranking with rule-based constraint checking.
Presents MatchBERT, a fine-tuned cross-encoder for clinical trial eligibility scoring. Evaluated against TREC 2021 Clinical Trials Track gold labels, achieving NDCG@10 of 0.900 on internal silver-label validation and AUC 0.703 on the external gold-label test set — a 49% improvement over the BM25 baseline.
MODEL CARD
A domain-driven deep learning model for clinical trial eligibility scoring, built on a model- and platform-agnostic retrieve → rank → explain pipeline. Trained on 87.8M patient-trial pairs with validated silver labels.
0.703
AUC, external gold labels
87.8M
Training pairs
3 sec
Per patient
520K+
Trials indexed
matchvox.ai/research
Evaluated against TREC 2021 Clinical Trials Track gold-standard expert labels. Not self-reported metrics on hand-picked benchmarks.
AUC 0.703 on external test set
We use multiple independent explanation methods. When methods disagree, clinicians are told — not shown cherry-picked explanations.
Multiple independent methods compared
Neural ranking for semantic understanding + rule-based constraints for boolean eligibility criteria. Best of both worlds.
5,418× faster than exhaustive pairwise scoring
Create a free account and run a match in under 2 minutes.