Uncovering interpretable potential confounders in electronic medical records

Abstract

In medicine, randomized clinical trials (RCT) are the gold standard for informing treatment decisions. Observational comparative effectiveness research (CER) is often plagued by selection bias, and expert-selected covariates may not be sufficient to adjust for confounding. We explore how the unstructured clinical text in electronic medical records (EMR) can be used to reduce selection bias and improve medical practice. We develop a method based on natural language processing to uncover interpretable potential confounders from the clinical text. We validate our method by comparing the hazard ratio (HR) from survival analysis with and without the confounders against the results from established RCTs. We apply our method to four study cohorts built from localized prostate and lung cancer datasets from the Stanford Cancer Institute Research Database and show that our method adjusts the HR estimate towards the RCT results. We further confirm that the uncovered terms can be interpreted by an oncologist as potential confounders. This research helps enable more credible causal inference using data from EMRs, offers a transparent way to improve the design of observational CER, and could inform high-stake medical decisions. Our method can also be applied to studies within and beyond medicine to extract important information from observational data to support decisions.

Natural Language Processing for Smart Healthcare

“…Before the interpretability issue is fully explored, the role of decision support systems in clinical practice can only be auxiliary from the perspectives of medical ethics and practical application. [133], [134] machine translation clinical documentation [135]- [137] speech recognition clinical decision support build QA-based clinical decision support systems [88], [138], [139] information extraction build clinical decision support systems with extracted information: family history information [140], entities and relations [141], [142], treatment and prognosis data [143], clinical data concepts and features [144], causal relations [113], [114] question answering healthcare quality control: assess clinical procedures [145], [146], warning of ADE [147], disease symptoms [148], [149], and outcome-related causal effects [150] information extraction, causal inference provide supporting evidence for decisions under evidence-based fashion [112], [151]- [154] information retrieval, causal inference Hospital Management medical resource allocation patient triage [155], [156] information extraction enable users to communicate and control intelligent systems through virtual assistants [157]- [159], hospital automation systems [160], [161] and collaborative robots [162], [163] speech recognition, natural language understanding predict and reduce readmission rate [164]- [166] information extraction free medical staff from routine text writing [74], [167] information extraction data management m...…”

“…Recently, with growing concerns about uninterpretable black box models, the importance of causal inference has gradually been recognized by NLP researchers, especially in the area of healthcare. Specifically, recent advances of causal inference in NLP for smart healthcare have been made in uncovering causality from medical text [112]- [114] and realizing reliable NLP-driven applications with discovered causal effects [112]- [114].…”

“…In addition, to ensure quality control and future quality improvement, NLP can also contribute to the assessment of clinical procedures [145], [146], warning of potentially harmful adverse drug events (ADEs) [147], disease symptoms [148], [149], and outcome-related causal effects [150]. Finally, NLP is also powerful in enhancing the interpretability and reliability of clinical decision support systems for practical deployment, by providing supporting evidence for diagnosis or treatment decisions in an evidencebased fashion [112], [151]- [154].…”