Before you can extract anything useful from a clinical note, you need to understand what kind of document you are actually reading. The SOAP format (Subjective, Objective, Assessment, Plan) is taught in medical school and referenced in EHR documentation standards, but it describes an ideal that practicing clinicians follow loosely, if at all. The structure of a real progress note is shaped by documentation habits, EHR template design, time pressure, and specialty convention. None of those factors stay consistent across a patient population.
When we built CliniNote, we spent considerable time just characterizing the structural variation in the note corpora we were working with. Two attending physicians at the same hospital, in the same department, documenting the same type of encounter, could produce notes with almost no shared structural features beyond the patient name header. That variation is not a bug in how clinicians document; it is a reflection of how clinical communication works. But it is a significant challenge for any extraction system that assumes consistent structure.
What SOAP Actually Looks Like in Practice
In an ideal SOAP note, the Subjective section contains what the patient reports: chief complaint, history of present illness, symptom characteristics, and relevant history. The Objective section captures what the clinician observes and measures: vital signs, physical examination findings, and relevant lab or imaging results. The Assessment summarizes the clinician's interpretation, typically as a problem list or diagnostic impression. The Plan documents the intended management: medications, referrals, follow-up, and orders.
In practice, outpatient notes often collapse Assessment and Plan into a single A/P block. Inpatient attending notes may separate them clearly, while resident notes use templates that front-load vitals and lab results into the Objective section with minimal narrative. Nursing notes follow entirely different conventions, often structured around care interventions rather than diagnostic reasoning. Procedural notes document what was done rather than why, with the indication as a brief header rather than a developed section.
The section header itself, when present, is one signal. But headers are often absent, abbreviated, or inconsistently applied. A note labeled "SOAP note" may have the Assessment content embedded in a long HPI narrative rather than in a distinct section. Recognizing section boundaries requires contextual signals, not just header matching.
The Subjective Section: Where Context Lives
The subjective section is the richest source of temporal context and patient-reported symptoms. The history of present illness (HPI) is typically written as a narrative: "Patient is a 64-year-old male with known hypertension and type 2 diabetes who presents with three days of worsening shortness of breath." From an extraction standpoint, this single sentence contains a demographic anchor, two conditions stated as background history, the chief complaint, and a duration.
The challenge is that the HPI also contains conditions the patient does not currently have, conditions from the past, and conditions the clinician is ruling out. "He denies chest pain or orthopnea" sits in the same section as "he reports bilateral ankle swelling." Both involve the word "pain" or condition references, but they carry opposite assertion statuses. The subjective section is where negation handling matters most, and where general-purpose extraction fails most visibly.
There is also the problem of chronological mixing. An HPI written in the present tense may describe a symptom that started six months ago. The note does not flag this explicitly. The phrase "he has been managing his heart failure with furosemide for the past two years" introduces a historical exposure that a research extraction system needs to date correctly, not treat as a new finding.
Assessment and Plan: The Most Extraction-Relevant Sections
For most RWE and research extraction tasks, the Assessment section carries the highest signal density. This is where the clinician states their diagnostic conclusions: not as a narrative but typically as a problem list, with each item labeled as active, chronic, or new. Diagnoses in the Assessment are the clinician's committed conclusions. The same condition mentioned in the HPI as patient-reported is distinct in research value from the same condition appearing in the Assessment as a clinician-confirmed diagnosis.
In ICD-coding workflows, coders are trained to assign codes based on conditions documented in the Assessment, not from HPI references. Research extraction should follow similar logic. A condition mentioned only in the HPI ("history of prior MI noted by patient") is not equivalent to a condition listed under Assessment as active ("Ischemic cardiomyopathy, stable"). The section where an entity appears shapes what it means.
The Plan section is critical for medication extraction. New prescriptions, dose changes, and discontinuations are typically documented here, either as free text or as order references. "Continue lisinopril 10mg daily, increase furosemide to 40mg BID" encodes a persistence instruction and a dose change in a single sentence. For pharmacoepidemiological studies, this type of information cannot be reliably captured from prescription claims data alone, which records dispensing events but not the clinical rationale or dose adjustment intent.
Section Segmentation and Why It Matters for Extraction
Section segmentation is the process of classifying each sentence or paragraph in a note into its corresponding section type, even when explicit headers are absent. This is a prerequisite for high-accuracy entity extraction in most clinical NLP systems, including ours.
Consider a note where the same medication appears in three places: once in the medication reconciliation list, once in the HPI as a medication the patient ran out of, and once in the Plan as a newly adjusted dose. A system that extracts medication mentions without section context will aggregate all three as equivalent. A system that knows which section each mention belongs to can correctly identify that the first is a historical record, the second is an exposure gap, and the third is the current clinical action.
We trained a section classifier on a corpus of labeled notes from multiple specialty types, including internal medicine, cardiology, and nephrology outpatient encounters. The classifier assigns section labels at the sentence level and handles unlabeled sections by propagating the label from the nearest preceding header combined with linguistic features. On sections with explicit headers, accuracy is above 96%. On headerless notes, accuracy drops to roughly 88% for Assessment content and 85% for Plan content, which is why we treat section-ambiguous extractions with lower default confidence scores.
What CliniNote Does When Structure Is Ambiguous
Not every note can be segmented reliably. Very short notes, procedure notes, and notes written as a single undivided paragraph represent the hardest cases. For these, we do not attempt to assign section labels at the sentence level. Instead, we flag the note as structurally ambiguous and apply entity-level confidence adjustments that reflect the higher uncertainty about each entity's assertion status and clinical role.
We are not claiming that this solves the problem. A structurally ambiguous note produces lower-confidence extractions, and in some research contexts, those notes should be excluded from automated extraction and routed to human review. That is the honest answer. Trying to over-segment a note that genuinely has no discernible structure does not improve accuracy; it produces false confidence in outputs that look correct but are not.
What understanding note anatomy does is shift the extraction problem from "find all mentions of condition X" to "find clinician-confirmed, currently-active instances of condition X in the Assessment section, with a note date that falls within the study window." That is a harder problem with much higher research value, and it requires knowing what section you are reading.
