For endocrinologists, keeping pace with an expanding body of evidence while managing increasingly complex patient histories is a familiar challenge. Natural language processing (NLP) is beginning to change how that information enters everyday clinical practice. 

 As an endocrinologist who has also co-founded health technology and medical education technology companies, I have seen that change firsthand. NLP has made me a better clinician, largely by making current evidence usable at the point of care rather than months later. It has made me more efficient and a clearer communicator with patients and referring colleagues, because the full longitudinal picture is in front of me instead of reconstructed from memory. It has made me more confident in my decisions and has meaningfully reduced my burnout. Here’s a look at how these tools are being used today, and where this technology can take clinical care next. 

 The language of NLPs 

NLP is the branch of AI that enables computers to read human language and extract meaning from it. Not structured data, not lab values and ICD codes, but the actual language clinicians write: the progress notes, the radiology reports, the pathology summaries, the referral letters. The text that contains most of what we know about a patient, and that no traditional analytics tool has ever been able to read. 

 What makes the capabilities in these examples possible is not NLP alone. NLP provides the foundational layer by identifying entities, such as drug names, gene variants, and lab values, recognizing relationships between them, and linking clinical concepts to standardized ontologies. Large language models (LLMs) then apply inferencing on top of that structured output.  

 The defining strength of modern LLMs is that many of these inferencing capabilities are available out of the box, without requiring a separate AI system to be built for each clinical task. A single model can cross-reference a medication against a procedure note, reconstruct a lab trajectory, compare a pathology finding against an updated guideline, and draft an appeal letter. This convergence of NLP's extraction power with LLM reasoning is what clinicians are beginning to encounter in practice, even when the underlying architecture goes unexplained. 

 NLPs in everyday practice

The value of NLP becomes clearest when you look at what it can do in the day-to-day reality of clinical practice. The three examples that follow come from my own practice—situations in which NLP helped, could have helped in hindsight, or may soon be able to help. Each one illustrates a capability that clinical work depends on, and that traditional tools cannot provide. Together, they describe what I think is the central problem in modern clinical care: the human clinician has a limited context window. We can only hold so much information at once. NLP does not have that constraint. 

 How NLP helped: When guidelines change and the chart doesn't know 

A 54-year-old woman undergoes total thyroidectomy for a 2.3 cm follicular variant papillary thyroid carcinoma (PTC). The pathology report describes an infiltrative growth pattern. Using the 2015 American Thyroid Association (ATA) guidelines, the framework I had relied on for years, I risk-stratify her as low. Radioactive iodine (RAI) is deferred. Surveillance begins. 

 What I did not yet know: the 2025 ATA guidelines replaced the prior three-tiered risk stratification system with a four-tiered system (low, low-intermediate, intermediate-high, high) that incorporates refined pathologic features.  

Under this updated composite assessment, a 2.3 cm infiltrative follicular variant PTC, which would have been classified as low risk under the 2015 system, may now meet criteria for the new low-intermediate risk category depending on the full constellation of pathologic features (for example, vascular invasion, nodal status, and molecular profile).  

 This reclassification, from low to low-intermediate risk, directly affects the RAI decision framework. The guidelines are long, and practice-changing updates are not labeled as such. I found this one only because I ran the case past an open-access, HIPAA-compliant clinical decision support tool as a second opinion, which flagged the reclassification against her pathology report and made the correct recommendation under the updated guidelines. Guideline adoption has been documented to lag publication for years; for this patient, that lag was a single clinic visit. It closed only because I thought to ask. 

NLP's role:NLP performed literature integration, continuously ingesting new guideline publications and cross-referencing them against the language already in her pathology report.  NLP integrated directly into the electronic health record is not currently available. But NLP-LLM integration could soon make it possible to run that same check inside the electronic health record and surface the discrepancy at the point of care for every patient rather than only the ones a clinician chooses to second-guess.  

 How NLP could have helped: Connecting the diagnosis dots 

A patient's chart may contain all the clues to a diagnosis, even when no single clinician has seen enough of them to make the connection. Case in point: A 71-year-old man with HIV, well-controlled on Symtuza (fixed-dose combination of darunavir, cobicistat, emtricitabine, and tenofovir alafenamide), is referred for new-onset Cushingoid features: central weight gain, moon facies, easy bruising, proximal weakness. His infectious disease physician suspects antiretroviral lipodystrophy. His primary care physician considers alcohol use. The endocrinology referral simply says: "rule out Cushing's." 

 Twenty minutes into the chart review, the picture forms, but only because I happened to read across every department simultaneously. Orthopedics had administered three intra-articular triamcinolone injections over 6 months for bilateral knee osteoarthritis. The HIV clinic continued Symtuza throughout without change. Nobody had flagged the interaction. Cobicistat, the pharmacokinetic booster in Symtuza, is a strong CYP3A4 inhibitor. It dramatically increases systemic steroid exposure from injections that clinicians and patients assume are purely local. The result was exogenous Cushing's syndrome, HPA axis suppression, and a patient at real risk for adrenal crisis the moment the injections stopped. 

 Every clue existed in the chart. What did not exist was any mechanism to assemble them. The orthopedic procedure note, the HIV clinic medication list, and the endocrine referral were each accurate and each incomplete, and the system that stored all three was never designed to read across them. Busy clinicians were asked to do that work manually, in the margins of a full clinic day, while the connection sat in the record for 6 months. 

 NLP's role:NLP could have generated the alert earlier, before the HPA axis was fully suppressed. It could have performed relation extraction, linking cobicistat in the medication list to corticosteroid injections in the procedure notes across providers.  

 How NLP may soon help: Clinical text summarization 

Here's a classic example of the administrative wall endocrinologists know too well. A 72-year-old woman with osteoporosis sustains her second vertebral fracture while on IV zoledronic acid (Reclast). She has severe osteoporosis, documented progression despite antiresorptive therapy, and has no cardiovascular contraindication. Romosozumab is clinically indicated. The prior authorization request is denied within 48 hours. The fracture history, medication timeline, BMD trajectory, and cardiovascular documentation all exist in the chart. The appeal letter takes 15 minutes to write. 

 NLP's role: Using NLP can convert those 15 minutes of administrative work into a review-and-sign task, returning that time to the clinical encounter where it belongs. NLP can summarize clinical text, read the longitudinal record, extract the clinically relevant findings, map them against payer-specific authorization criteria, and generate a structured draft appeal letter.  

 What NLP cannot do 

None of these examples are arguments for replacing clinical judgment, and the limitations here deserve more than a passing disclaimer. 

 NLP, as an extraction layer, makes well-characterized errors. It misreads negation, struggles with ambiguous abbreviations, and underperforms outside its training domain. These are manageable risks that validation can quantify, and monitoring can catch. 

 The risks introduced by LLM inferencing are different and more insidious. LLMs generate conclusions that are fluent, confident, and occasionally wrong in ways that are difficult to detect without independent verification. A clinician who receives an AI-generated insight, particularly one that fits a plausible narrative, may anchor to it without critically re-examining the underlying data. This is not a hypothetical failure mode. Automation bias, the tendency to accept machine outputs uncritically, is a documented phenomenon in clinical settings, and LLMs are unusually well suited to triggering it because their outputs read as reasoned and authoritative rather than mechanical. 

 The mitigation is not to avoid these tools but to use them with discipline. AI-generated insights should be treated as a starting point for clinical reasoning, not a conclusion. Outputs should be traceable to the source notes that generated them. Institutions deploying these systems should require that the model shows its work, not just its answer. Clinicians should remain alert to the possibility that a coherent AI-generated narrative confidently describes something that is not true. 

 What these clinical cases do argue for is this: the volume of information in a modern patient's chart has outgrown the human context window. No clinician, however skilled or diligent, in the same afternoon, can simultaneously read across cardiology, oncology, orthopedics, genetics, and pharmacy, plot a longitudinal lab trajectory, cross-reference a pathology report against a guideline published last month, and draft an appeal letter, all for the same patient. 

 NLP does not replace what endocrinologists do. It reads the parts of the chart we do not have time to reach, and surfaces what it finds before the moment to act has passed. Used with discipline, that is a meaningful addition to clinical practice. Used carelessly, it adds a new class of confident, well-written errors to a chart that already has too many. 

Chaitanya Mamillapalli, MD, MRCP, FAPCR, is Chief of Endocrinology at Springfield Clinic in Springfield, Illinois, and an adjunct faculty member at Southern Illinois University School of Medicine and the University of North Dakota School of Medicine. He is founder and CEO of CareHealth.ai, a health technology company, and MedEdX.ai, an adaptive medical education platform that uses Socratic, case-based learning to develop clinical judgment. 

 Acknowledgment: The author thanks Senthil Nachimuthu, MD, PhD, for his editorial input. Dr. Nachimuthu is a physician informatician, physician-scientist, faculty member at the University of Utah and the VA in Salt Lake City, and a former health technology executive.