
Clinical notes hold the story behind every lab result and prescription, but much of it sits in handwriting, scans and free text. This guide covers the 8 things that decide whether structured notes can be trusted by the clinicians, systems and AI tools that rely on them.
Lab results tell you what was measured. Clinical notes tell you why. Discharge summaries, referral letters, consult notes and ward notes carry the reasoning, the history, the doubts and the decisions that structured fields rarely capture. They are also where a great deal of clinical information still lives on paper or in scanned PDFs, particularly in hospitals and clinics that digitised recently or are still in transition.
Turning these notes into structured data, usually FHIR (Fast Healthcare Interoperability Resources), is now a common requirement. Electronic medical record (EMR) vendors need it to bring historical records into their systems. Digital health and AI teams need it to give their products clinical context. Hospitals and registries need it to make years of paper records usable. The difficulty is that reading the words is only the beginning. A note can say that a disease is absent, suspected, historical or belongs to a relative, and a careless conversion will record all of these as confirmed diagnoses.
This guide is written for EMR and clinic system vendors, digital health and AI product teams, and hospitals and registries digitising records. It is organised in three parts: understanding the note, interpreting what it says, and producing output you can trust.
Figure 1. Six short lines from a typical note, and the six structured facts they should become. Only two of them describe something the patient has right now. The rest are absent, suspected, about a relative or no longer true, and each status matters as much as the term itself.
Figure 1 is the heart of this guide. If a conversion captures only the terms on the left (breathlessness, chest pain, chest infection, diabetes, metformin, amlodipine), the patient appears to have four diagnoses and two current medicines. The real picture is very different, and the 8 points below are about getting it right.
The right level of structure depends on what the data is for. If the goal is continuity of care, a clinician may only need the note searchable, filed against the right patient and split into sections such as history, examination and plan. If the goal is coding, analytics, a registry or an AI product, the diagnoses, medications, allergies and key findings need to become discrete, coded data with the right status attached.
Writing down the purpose first saves a great deal of rework. It tells you which elements must be extracted, which can stay as narrative, how much clinical review is justified, and which output format the receiving system expects. It also stops a common pitfall: extracting everything at maximum depth, at maximum cost, when the use case only needed a fraction of it.
"Clinical notes" covers documents with very different shapes. Table 1 summarises the main types and what each tends to contain.
Table 1. Common clinical note types, their usual structure and the points that most often trip up a conversion.
As Table 1 shows, a single extraction approach rarely fits all types. Identify the note type first, then apply rules suited to it. Research on clinical text has long used the fact that discharge summaries and admission notes are divided into sections, such as the chief complaint, to help interpret what the text means.4 The same word can carry different weight in a history section, an assessment and a plan.
Handwriting is the obvious challenge. In one study of case notes from a Spanish general hospital, 15 per cent of clinical histories were so illegible that the meaning was unclear.5 A 2025 case study on digitising historical handwritten clinical records concluded that general-purpose OCR (optical character recognition) systems were insufficient for reliable clinical data extraction, because of the complexity and variability of handwritten medical records.6
Abbreviations are the less obvious challenge, and they affect typed notes too. One study found that 54.3 per cent of abbreviations in clinical notes were ambiguous, and that the same abbreviation can mean different things in different institutions, specialties, between practitioners and even within the same note.3 In Figure 1, "SOB" means shortness of breath, "2/7" means two days and "bd" means twice daily, but each of these conventions varies by country, specialty and hospital. Build a list of local abbreviations with each client or site, and make sure anything ambiguous is flagged for clinical review rather than guessed.
Clinical notes are full of findings that are documented precisely because they are not present. Researchers have studied this for more than two decades: NegEx, an early and widely used method for detecting negated findings, was developed because so many findings and diseases in discharge summaries are recorded as absent.1 Later work extended the idea to other contextual features, including whether a finding is historical or relates to someone other than the patient.2
Four patterns deserve explicit rules. Negation ("no chest pain", "denies fever"). Uncertainty ("?LRTI" for a suspected lower respiratory tract infection, "query TB" for suspected tuberculosis, "possible", "rule out"). Family history ("mother: DM2" for a mother with type 2 diabetes, "FHx of breast cancer"). And hypotheticals or instructions ("return if chest pain develops"). FHIR gives you the means to represent these properly: the Condition resource has a verificationStatus element with values including unconfirmed, provisional, differential, confirmed and refuted, and family history belongs in its own resource rather than the patient's problem list.7 The risk is not that the standard cannot express these states. It is that a conversion never looks for them.
Notes rarely give tidy dates. They say "2/7", "since last week", "post-op day 3", "previous MI" (a past myocardial infarction) or "as before". Turning these into dates requires knowing the date of the note itself, the date of the encounter and the conventions of the writer. Researchers have found temporal information in clinical text harder to extract than in general English text, partly because clinical writing is so informal.8
Three distinctions matter most: when something started (onset), whether it is current or historical, and whether it has resolved. In FHIR, a Condition carries a clinicalStatus (for example active, remission or resolved) separately from its verification status, along with onset and abatement dates.7 Where a note only gives a relative time, record how the date was derived, so nobody later mistakes an estimate for an exact date.
Medication lines in notes combine several hard problems at once: brand and generic names, local brands that differ by country, dose and frequency shorthand ("500 bd", "1-0-1", "nocte"), and changes over time. A discharge summary may list medicines that were started, changed, held or stopped during the admission, and these must not all become a single "current medications" list.
FHIR's MedicationStatement resource requires a status, with values including active, completed, intended, stopped, on-hold and not-taken.9 Capture the medicine, strength, dose, route, frequency and status as separate elements, keep the original wording, and flag anything where the status cannot be determined from the note. In Figure 1, metformin and amlodipine look similar on the page but carry opposite meaning for anyone reconciling the patient's medicines.
A structured note has two layers: the document and the facts inside it. In FHIR, the document is usually represented as a Composition (or referenced through a DocumentReference), which is organised into sections, each containing narrative text, structured entries or both.10 The facts become their own resources, coded to medical coding standards so that systems can use them.
Figure 2. A structured clinical note has two layers. The document layer keeps the note's sections and original narrative so a clinician can read it as written. The fact layer turns what the note says into coded resources, each with the status that tells a system whether it is current, suspected, ruled out, stopped or about a relative.
Figure 2 shows why keeping the original narrative matters. Structured facts are what systems and AI tools use, but clinicians, auditors and reviewers still need to read the note as it was written. Keeping both layers also makes it possible to improve the structuring later without re-digitising the source. Where a client or institution has its own code sets, map to those as well as to medical coding standards, so the output fits the receiving system without forcing it to change.
No automated approach will interpret every note correctly, so the design question is where clinically trained reviewers step in. The most practical pattern is flag-based review: the system structures the note, checks flag anything uncertain (an ambiguous abbreviation, an unreadable dose, a status that cannot be determined), and reviewers resolve those items rather than re-reading every note. Agree with the receiving team what triggers review and what happens to items that remain unresolved.
Privacy needs particular care with free text, because identifiers are not confined to labelled fields. A note can mention the patient's name, a relative's name, a ward, a date of birth or a phone number anywhere in the narrative. A study of automated de-identification found that fully customised systems removed 97 to 99 per cent of personally identifying information, but also that free-text notes vary widely with note purpose and institutional conventions, and that no system can guarantee performance on every note it will meet.11 If notes will be used beyond direct care, test pseudonymisation on your own notes, decide what level of residual risk is acceptable, and keep governance around who can access the original text.
Table 2 brings the 8 points together so you can use them with your product, data and clinical teams.
Table 2. A checklist summarising the 8 things to check before structuring clinical notes.
The thread running through all 8 points: the term is only half the fact. Whether something is present, absent, suspected, historical, stopped or about someone else is what makes structured notes safe to use.
Teams usually come to us at one of two moments. Some are still deciding what to do with their notes: which note types matter, how deep the structure needs to go, and how to handle privacy. Others already know what they need and want notes converted reliably, at volume, into the format their system expects. We support both.
JondaX has processed more than 500,000 files and 26 million data points at 99.9 per cent field-level accuracy, across more than 10 languages. Clinical reviewer corrections feed back into future mapping, so results on your own notes improve as more are processed. Because note styles vary so much between sites, the most useful first step is often a sample run on a pseudonymised set of your own notes.
Yes, but reading the handwriting is only the first step. The text then has to be interpreted in clinical context (abbreviations, negation, uncertainty, timing and medications), mapped to the right FHIR resources and medical coding standards, and checked by clinically trained reviewers where the system is unsure.
General OCR turns images into text, but clinical notes are full of shorthand, ambiguous abbreviations and statements that say a condition is absent, suspected or belongs to a family member. A case study of handwritten clinical records found general-purpose OCR insufficient for reliable clinical data extraction,6 so notes need clinical interpretation on top of text recognition.
The note itself is usually represented as a Composition or DocumentReference, while the facts inside it become resources such as Condition, MedicationStatement, AllergyIntolerance, Observation and FamilyMemberHistory. Keeping the original narrative alongside the structured resources lets clinicians read the note as written.
Statements such as "no chest pain" or "?TB" must not become confirmed diagnoses. In FHIR, the Condition resource has a verificationStatus element with values such as provisional, differential, confirmed and refuted,7 so uncertainty and ruled-out findings can be represented rather than lost.
Names, dates, places and other identifiers can appear anywhere in a narrative, not just in labelled fields. Automated de-identification can remove most of them, but performance depends on note type and local conventions,11 so it should be tested on your own notes and combined with review and governance.
Yes. JondaX, from Jonda Health, has a Clinical Notes module (currently in Beta) that takes clinical notes from PDF or paper and returns structured notes in FHIR JSON and other formats, with clinical review built in. Email hello@jonda.health to try it.
About the author. Suhina Singh is the Founder & CEO of Jonda Health, a health data infrastructure company that turns fragmented clinical data, digital and non-digital, into structured, standardised information that clinicians and systems can actually use. Jonda Health is ISO 27001 certified and works with hospitals, laboratories, EMR vendors, digital health companies, registries and life sciences teams. Get in touch at hello@jonda.health.