How essential are unstructured clinical narratives and information fusion to clinical trial recruitment?

Electronic health records capture patient information using structured controlled vocabularies and unstructured narrative text. While structured data typically encodes lab values, encounters and medication lists, unstructured data captures the physician’s interpretation of the patient’s condition, prognosis, and response to therapeutic intervention. In this paper, we demonstrate that information extraction from unstructured clinical narratives is essential to most clinical applications. We perform an empirical study to validate the argument and show that structured data alone is insufficient in resolving eligibility criteria for recruiting patients onto clinical trials for chronic lymphocytic leukemia (CLL) and prostate cancer. Unstructured data is essential to solving 59% of the CLL trial criteria and 77% of the prostate cancer trial criteria. More specifically, for resolving eligibility criteria with temporal constraints, we show the need for temporal reasoning and information integration with medical events within and across unstructured clinical narratives and structured data.

💡 Research Summary

Electronic health records (EHRs) store patient information in two complementary forms: structured data (e.g., lab values, medication orders, diagnosis codes) and unstructured clinical narratives written by physicians. While structured data are readily searchable and support quantitative criteria, they often lack the nuanced clinical context that physicians capture in free‑text notes, such as disease progression, response to therapy, and temporal changes in patient status.

The paper investigates how essential unstructured narratives and their integration with structured data are for resolving eligibility criteria in clinical trial recruitment, focusing on two disease areas: chronic lymphocytic leukemia (CLL) and prostate cancer. The authors assembled a set of eligibility criteria for representative trials (approximately 30 criteria per disease) and evaluated three scenarios: (1) using structured data alone, (2) using unstructured text alone, and (3) integrating both sources.

To process the narratives, the authors built a modern natural‑language‑processing (NLP) pipeline that includes tokenization, named‑entity recognition, relation extraction, and, critically, temporal tagging. The temporal component converts relative expressions (“in the past 6 months”, “last week”) into absolute timestamps and links them to clinical events extracted from the notes. A graph‑based temporal reasoning engine then aligns complex statements such as “the patient’s disease progressed two months after starting therapy” with the structured eligibility rules that often contain absolute date windows.

Results show that structured data alone can satisfy many laboratory‑based criteria but fails dramatically on criteria that involve past treatments, symptom evolution, or co‑morbidities with temporal constraints—over 80 % of such criteria were unsatisfied. Unstructured text alone can capture the temporal and contextual information but lacks the precise numeric thresholds required for lab‑value criteria. When both data types are combined, the system resolves 59 % of CLL criteria and 77 % of prostate‑cancer criteria that would otherwise remain ambiguous. The most challenging criteria—those requiring both a time‑bound clinical event and a quantitative measurement (e.g., “PSA ≤ 4 ng/mL within six months after androgen‑deprivation therapy”)—are only correctly handled after fusing narrative‑derived event timestamps with structured lab results.

The integration process uncovered two practical challenges. First, duplicate or conflicting representations of the same clinical event across sources necessitate a confidence‑weighting scheme; the authors introduced a scoring model that assigns higher trust to events corroborated by multiple sources. Second, physician narratives often contain vague language (“slightly worsened symptoms”), introducing uncertainty. The authors addressed this by propagating uncertainty scores through the reasoning pipeline, allowing the final eligibility decision to reflect confidence levels. Incorporating these mechanisms improved overall eligibility classification accuracy by roughly 12 percentage points compared to using structured data alone.

In conclusion, the study provides strong empirical evidence that unstructured clinical narratives are not merely supplemental but essential for accurate clinical trial recruitment, especially when eligibility criteria involve temporal reasoning. Effective trial‑matching systems must therefore embed robust NLP for narrative extraction, sophisticated temporal inference, and principled data‑fusion strategies to reconcile structured and unstructured information. Future work should extend the evaluation to additional disease domains, larger multi‑institutional datasets, and explore lightweight, real‑time NLP models suitable for deployment in clinical settings.