What is clinical data quality?

How far a patient's record can be trusted for the job at hand: whether the facts that should be there are there, whether they are right, whether they agree with each other and with other sources, whether they make sense, and whether they are up to date.

How do you measure the quality of EHR data?

Pick the fields that matter for the use you have in mind, then measure how often they are present, how often they follow the expected format and code set, how often values fall outside plausible ranges, how many patients are duplicated, and how a sample of coded diagnoses compares with the case notes.

How accurate is hospital diagnosis coding?

It varies widely. A systematic review of 32 UK studies found a median accuracy of 80.3% for diagnosis coding against the case notes, ranging from 50% to 98% between studies. Coding accuracy improves when someone depends on it and checks it.

Why does data quality matter for AI?

AI learns patterns from records and acts on what it reads. Missing or wrong data produces confident wrong answers, and a model built on one hospital's data can fail on another's. Clean, standard, well-documented data is what makes a tool work and lets you check that it does.

Should data be cleaned after it is entered or at the point of entry?

At the point of entry wherever possible. Structured fields, pick lists, range checks and duplicate checks at registration prevent errors that are expensive or impossible to fix later, because by then nobody remembers what the right value was.

Clinical data quality: why AI in healthcare starts with clean records

Every promise about AI in healthcare rests on an assumption that is rarely said out loud: that the records it reads are right. A scribe that drafts a note, a model that flags a deteriorating patient, a study that compares two treatments across thousands of patients: each one is only as good as the data underneath it.

In most hospitals, that data is worse than anyone thinks, and nobody has measured it. This article explains what clinical data quality means, how to measure it, where it goes wrong in Indian hospitals and clinics, and how to fix it at the source.

What “data quality” means for clinical records

Researchers who study electronic health record (EHR) data have tried to pin the idea down. A widely cited review by Weiskopf and Weng in 2013 looked at 95 studies, found 27 different terms in use, and reduced them to five dimensions:

Completeness: is a fact about the patient that should be in the record actually there?
Correctness: is what is recorded true?
Concordance: do parts of the record agree with each other, and with other sources?
Plausibility: does a value make sense given everything else we know?
Currency: does the record show the patient’s state at the time it claims to?

A later framework from Kahn and colleagues in 2016 grouped checks into three kinds, which map neatly onto tests you can actually run:

Conformance: do values follow the expected format, code set and relationships? A date that is a date, a diagnosis that is a valid ICD-10 code, a lab result attached to an order that exists.
Completeness: how often is a field filled in at all?
Plausibility: are values believable? A pulse of 720, a discharge before the admission, a man recorded as pregnant.

The same paper separates verification (checking data against your own rules and expectations) from validation (checking it against an outside reference, such as the paper case notes or another system). Both are needed. Verification is cheap and automatic; validation is slower and catches what rules cannot.

Where clinical data goes wrong

Most of the problems are mundane, and most happen at the moment of entry.

Paper and free text. A diagnosis written in a sentence, a prescription photographed, a discharge summary scanned. Readable by a person, invisible to a query.

The same patient twice. A patient registers as “S. Kumar” in March and “Suresh Kumar” in June, with a new phone number. Their history is split across two records, and anyone looking at either sees half of it.

Units and decimal slips. A weight in pounds in a kilogram field, a lens power of 250 when 2.50 was meant, a haemoglobin entered as 115 when the lab reports 11.5 g/dL.

Loose coding. The right ICD-10 code takes time to find, so a vaguer one is chosen, or none. A systematic review of 32 UK studies found a median accuracy of 80.3% for diagnosis coding when checked against the case notes, with individual studies ranging from 50.5% to 97.8%. Accuracy rose sharply after hospital payments came to depend on it, which says something about what makes coding improve. We know of no equivalent published figure for discharge coding in Indian hospitals, but a 2025 study at a North Indian tertiary hospital found that only about 1 in 10 of 1,147 cause-of-death certificates had been filled in correctly.

Copy-forward. Yesterday’s examination carried into today’s note because it was quicker than writing a new one. The record looks complete and is out of date.

Missing negatives. “No known allergies” and “allergies not asked” look the same when the field is simply empty.

How to measure it

You cannot fix what you have not measured, and you do not need a data warehouse to start. Pick the use you care about, then the handful of fields that use depends on, and measure:

Completeness: the share of visits with each key field filled: allergies, vitals, diagnosis, medicines, the follow-up date.
Conformance: the share of diagnoses that are valid ICD-10 codes, lab tests mapped to a standard code, dates in range, records that pass a FHIR validator.
Plausibility: values outside clinical ranges, impossible sequences of dates, contradictions such as a diagnosis on a patient of the wrong sex.
Duplicates: patients sharing a phone number and a similar name, or the same name and date of birth.
Coding accuracy: a sample of coded diagnoses checked against the notes by someone who knows the specialty.
Timeliness: how long after the visit the record is complete.

Do it once to get a baseline, then repeat it on a schedule. A number that is never measured again does not improve.

Fix it at the source

Cleaning data after the fact is slow, and often impossible, because by then nobody remembers what the right value was. The cheaper fix is to make the right entry the easy one:

Structured fields for what will be counted later: diagnoses, medicines, vitals, results.
Pick lists and search for diagnoses and medicines, mapped to standard codes underneath, so the doctor chooses “type 2 diabetes” and the record holds the code.
Range checks that warn, not block. A value outside the normal range should be flagged, and a likely decimal slip should be suggested, but a genuinely extreme value must still be recordable. A check that has to be defeated to record an emergency gets defeated by habit.
A duplicate check at registration. When a phone number already belongs to a patient, ask whether this is the same person before creating a new record.
Explicit negatives. “No known allergies” as a choice, distinct from a blank.
Corrections that leave a trail. When a record is wrong, correct it with a reason and keep the original, rather than overwriting it. That trail is what lets you trust the rest.
Standards at the edges. Lab results coded to LOINC, data exchanged as FHIR, and validated before it leaves.

Data quality and AI

AI adds two problems of its own.

The first is that a model inherits its training data. If certain patients are under-recorded, or certain diagnoses are coded loosely, the model learns those gaps.

The second is that a model built on one hospital’s data can fail on another’s. A widely cited example is a sepsis prediction model used in many US hospitals: checked at Michigan Medicine against 38,455 hospital stays, it missed 67% of sepsis cases while alerting on 18% of all patients. Data quality is part of why: differences in how and when things are recorded change what a model sees.

So before relying on any AI tool, ask what data it was built and tested on, and test it on a sample of your own records. And keep measuring after go-live, because recording habits change.

A data quality checklist

Have you chosen the fields that matter for the use you have in mind?
Do you know how complete each of them is today?
Are diagnoses coded in ICD-10, and lab tests mapped to LOINC?
Do range checks warn on implausible values, and suggest likely decimal slips?
Is there a duplicate-patient check at registration?
Can “none” be recorded explicitly, separately from “not asked”?
Are corrections made with a reason, keeping the original?
Has someone checked a sample of coded diagnoses against the notes this year?
Is data validated (for example against FHIR) before it is shared?
Do you repeat the measurements on a schedule, and does someone own the results?

Where 92health fits

92health is an EMR and AI platform used by clinics, hospitals, pharmacies, labs and optical shops in India. Much of the checklist above is built into how it records care:

diagnoses are chosen from searchable lists and stored with their ICD-10 codes;
on the eye work-up, a reading outside its normal range is flagged, and a likely decimal slip is suggested, without blocking a genuinely extreme value;
registering a phone number that already belongs to a patient asks whether it is the same person before a second record is made;
lab results are validated as FHIR R4 before they are shared;
a diagnosis recorded by mistake is withdrawn with a reason rather than deleted, and records entered later from paper are marked as such.

For hospitals that want to know where they stand, we run data quality reviews as part of our AI readiness work: a sample of your records, measured against the checks above, and a written list of what to fix first.

Sources

Weiskopf NG, Weng C. Methods and dimensions of electronic health record data quality assessment: enabling reuse for clinical research. JAMIA, 2013;20(1):144-151. doi:10.1136/amiajnl-2011-000681
Kahn MG et al. A harmonized data quality assessment terminology and framework for the secondary use of electronic health record data. eGEMs, 2016;4(1):1244. doi:10.13063/2327-9214.1244
Burns EM et al. Systematic review of discharge coding accuracy. Journal of Public Health, 2012;34(1):138-148. doi:10.1093/pubmed/fdr054
Srinivasan K et al. The final diagnosis: evaluating errors in cause-of-death reporting. Cureus, 2025;17:e82785. doi:10.7759/cureus.82785
Wong A et al. External validation of a widely implemented proprietary sepsis prediction model in hospitalized patients. JAMA Internal Medicine, 2021. doi:10.1001/jamainternmed.2021.2626

This article is general health information for readers in India. It is not a diagnosis, a prescription or a substitute for a consultation. Costs, procedures and regulations vary by clinic and state; confirm anything that matters with your own doctor or pharmacist.