Registry & Real-World Data Digest — Issue 02

Renee M Cholyway, William C Kethman

Summary: This review article surveys the ways electronic health records and AI are being used to support clinical decisions. Topics covered include early detection of sepsis, tools that reduce the time clinicians spend writing documentation, and systems that help prioritize patients who need attention. The authors note both the growing capabilities of these tools and real limitations — including the fact that AI systems are harder to train when data on a particular condition are scarce, and that the inner workings of many algorithms are not made transparent to the people using them. No new data are presented; this is a descriptive overview written for a surgical specialist audience.

Why it matters: For people new to these tools, the review brings together a broad landscape of AI applications into one readable summary. Its acknowledgment that data scarcity is a genuine obstacle — particularly for rare conditions — is an important point that enthusiastic accounts of AI often skip. Understanding this limitation matters for anyone involved in registries or research networks built around conditions that affect relatively small numbers of patients.

Expert perspective: This review covers familiar ground competently but does not critically examine whether the tools it describes actually work well in practice. Many AI clinical decision-support tools perform promisingly in controlled settings but show mixed results when used more broadly. Problems such as alert overload, disruption to clinical workflows, and the tendency of algorithms to perform poorly outside the populations they were originally trained on receive only passing attention here. The lack of a formal evidence-appraisal process means the review's conclusions reflect expert opinion rather than a systematic weighing of the available evidence. It is a reasonable starting point for someone new to the field, but not a definitive reference for those already working in health informatics or rare-disease data research.

Relevance: Clinical Informatics and Health Data Methods

Source: https://pubmed.ncbi.nlm.nih.gov/42553947/

---

Neha Simha, Hannah Takasuka, Li-Ching Chen, Umair Khan, Tomiko T Oskotsky, Marina Sirota, et al.

Summary: Researchers analyzed more than 640,000 clinical notes from nearly 26,000 pregnancies at a single academic medical center, using an automated computer program trained to recognize language that could stigmatize or demean patients. The program found that nearly half of all pregnancies had at least one note containing such language. Black patients and those with lower levels of education were significantly more likely to have stigmatizing language in their records. Pregnancies that ended in preterm birth were also more likely to contain this kind of language.

Why it matters: The words clinicians use in patient notes are not neutral. Language that demeans or makes assumptions about a patient can shape how future clinicians think about that person and influence the care they receive. The finding that Black patients are disproportionately described using stigmatizing language is consistent with well-documented inequities in obstetric outcomes and raises important questions about how documentation practices may contribute to those gaps. For health systems and researchers, this study shows that automated tools can be used to audit records for bias at a scale that was previously impossible with manual review.

Expert perspective: This study is a preprint — meaning it has not yet completed formal peer review — and the automated detection tool it uses has not been fully validated against a large, independently reviewed set of notes. Without knowing how often the tool incorrectly flags language as stigmatizing or misses language that is, the headline figure of 47% carries real uncertainty and should not be treated as a precise, established fact. The study also drew on one academic medical center, and documentation habits there may differ from those at community hospitals or safety-net clinics. It is also important to note that Black patients in this cohort may have had more complex medical situations or more clinical contact, which could partly explain the higher rates of flagged language — the study cannot fully rule out this possibility. The finding of more stigmatizing language in preterm births is similarly complicated by the fact that difficult pregnancies simply generate more documentation overall. Despite these cautions, the core finding — that measurable differences in language exist across racial groups and can be detected at scale — is credible and important. The practical implication, that health systems should be auditing their records for documentation bias, is sound.

Relevance: Real-World Data; Data Quality; Clinical Informatics

Source: https://pubmed.ncbi.nlm.nih.gov/42619892/

---

Sarah B Calvert

Summary: This essay, published in a clinical infectious diseases journal, raises a concern about AI tools that help clinicians write their notes more quickly. The author argues that while these tools reduce administrative burden, they may also produce notes that are more uniform and less detailed — stripping away the interpretive, patient-centered language that clinicians have traditionally used to capture the complexity and ambiguity of a patient's condition. Because patients in the United States have been able to read their own clinical notes since 2021, the quality and tone of that language now matters in a new way.

Why it matters: Clinical notes have always served two purposes: creating a legal and administrative record, and documenting a clinician's reasoning about a complex situation. In fields like infectious diseases, where uncertainty is common and conditions can carry significant social and psychological weight, the way a note is written can genuinely affect how future clinicians, consultants, and patients themselves understand the situation. If AI-generated notes smooth over uncertainty or use more generic language, something clinically meaningful may be lost — though how often this happens and whether it affects patient outcomes has not yet been studied rigorously.

Expert perspective: This is a thoughtful essay based on professional reasoning, not a research study with data. The concern it raises — that AI documentation tools may flatten the narrative quality of clinical notes — is plausible and has been echoed informally by some clinicians, but it has not yet been tested through systematic comparison of AI-generated and clinician-written notes. It is also worth noting that clinical notes before AI tools varied enormously in quality; many were already template-driven, repetitive, or generic. AI tools might in some cases improve on that baseline, particularly for clinicians who were not strong writers to begin with. The essay does not fully acknowledge this. The most useful takeaway is practical: health systems adopting AI documentation tools should build in quality standards, regular review of note content, and training that keeps clinicians responsible for the substance of what is recorded — rather than treating AI-generated drafts as finished documents.

Relevance: Clinical Informatics and Health Data Methods; Data Quality

Source: https://pubmed.ncbi.nlm.nih.gov/42617064/

---

Glenn K Goodrich, Elizabeth A Bayliss, James Lagrotteria, Jennifer C Barrow, Bill Harding, Kathy S Gleason, et al.

Summary: Researchers developed and tested a method to identify when a clinician deliberately decided to stop a patient's medication — as distinct from a patient simply not refilling a prescription or experiencing a gap in insurance coverage. Working with records from adults over 65 across five medication types, the team combined keyword searches of clinical notes with information from medication order records. When the two sources were used together, the method correctly identified intended discontinuations between 74% and 82% of the time, and correctly identified non-discontinuations between 75% and 95% of the time, depending on the medication type. Using either source alone performed worse.

Why it matters: Understanding whether a medication was deliberately stopped by a clinician — a practice called deprescribing — is important for studying whether reducing medications in older adults is beneficial or harmful. The problem is that standard insurance billing records, which researchers often rely on, cannot distinguish between a deliberate clinical decision and a patient simply not picking up a refill. If researchers cannot make that distinction reliably, studies of deprescribing can produce misleading results. This tool offers a practical step toward solving that problem using information already available in electronic health records.

Expert perspective: The approach is methodologically sound and the problem it addresses is real and underappreciated in medication research. However, there are important limitations to keep in mind. The tool was developed and tested at a single health system, so it may not work as well in settings with different documentation habits or different electronic record systems. The groups of records used to test each medication type were relatively small — as few as 80 for one category — which means the accuracy figures reported are estimates with meaningful uncertainty around them. A miss rate of roughly one in four to one in five intended discontinuations could meaningfully distort the conclusions of a study depending on what is being measured. Researchers who want to use this approach should consider testing it in their own setting before relying on it, and should think carefully about how any remaining errors might affect their particular analysis.

Relevance: Real-World Data; Real-World Evidence; Clinical Informatics; Data Quality; Observational Research

Source: https://pubmed.ncbi.nlm.nih.gov/42593400/

---

Joaquin M Espinosa, Kate Burdekin, Jeanette O Auman, Benjamin Tyndall, James F Cnota, Sigan L Hartley, et al.

Summary: This report summarizes a workshop organized by the NIH INCLUDE Project, which funds research into Down syndrome. Researchers from multiple institutions gathered to address the practical challenges of collecting and combining health data across many different study sites. The workshop reached four broad conclusions: using standardized definitions of data elements across sites is essential for combining data meaningfully; electronic health records can be used to feed research databases if extraction methods are carefully designed; AI tools, particularly those that read and interpret clinical text, can speed up the process of pulling information from records — but only if humans continue to review the results; and clear rules about privacy, fairness, and transparency must govern any AI use in this context.

Why it matters: The difficulties the INCLUDE Project encountered — inconsistent data definitions across sites, heavy manual workload, and the challenge of managing complex longitudinal data — are common to nearly every research network built around a rare or complex condition, including cystic fibrosis registries. The combination of standardized data definitions at the design stage with AI-assisted data extraction at the analysis stage, governed by clear oversight rules, represents a practical and increasingly mature approach to building better research infrastructure. The workshop's emphasis on governance — treating privacy, bias prevention, and transparency as operational requirements rather than aspirations — is particularly relevant as AI tools are adopted rapidly, and sometimes uncritically, in health data settings.

Expert perspective: This is a summary of workshop discussions, not a research study. There are no data, no measured outcomes, and no formal evaluation of whether the approaches described actually improve research quality. The recommendation that AI-assisted text analysis can accelerate data abstraction is consistent with broader evidence, but the specific performance characteristics, error rates, and costs in Down syndrome research or similar rare disease settings are not reported. The call for standardized data definitions is well-established in principle; the persistent challenge is getting institutions to adopt and maintain them over time, which workshop proceedings cannot guarantee. For those building or updating rare disease registries, the most practical takeaway is architectural: design data collection around shared standards from the beginning, use AI tools to assist rather than replace human review of clinical text, and build governance structures that are genuinely implemented rather than listed as good intentions. This report is a useful orientation, but it describes a direction of travel, not a proven destination.

Relevance: Patient Registries; Clinical Informatics; Data Standards; Data Governance; Real-World Data

Source: https://pubmed.ncbi.nlm.nih.gov/42586352/

---