The retrospective chart review is the workhorse of clinical research. It is fast, cheap, needs no recruitment, and produces genuinely useful findings about real practice.
It also produces the weakest Methods sections in medicine. The characteristic version reads: "We retrospectively reviewed the records of 84 patients who underwent X between 2020 and 2023. Data were collected and analysed using SPSS."
That paragraph is unassessable. A reader cannot tell how those 84 patients were found, how many were missed, who read the notes, what happened when the notes were ambiguous, or how much data was absent. And because none of it is visible, the reader has no choice but to discount the findings.
The remedy is not more data. It is ten specific things a Methods section must contain.
1. Name the design and the setting
State the design in the first sentence, then describe the setting concretely: how many centres, what type — tertiary referral, district general, teaching, private — which specialty units, and the catchment or case mix. Chart review findings are strongly setting-dependent, and a reader needs to know whether your setting resembles theirs.
We conducted a retrospective cohort study using medical records from two tertiary referral hospitals in southern India, both providing regional hepatobiliary services.
2. Define the study period, and say why
Exact dates, not years. And explain the boundaries: a protocol change, the arrival of a technique, the introduction of an electronic record system, or simply the period for which records are complete. Unexplained boundaries invite the suspicion that they were chosen after seeing the data.
3. Explain exactly how cases were identified
This is the item most often missing and the one that most determines credibility. How did you find your patients?
- ICD-10 codes? State every code used.
- Procedure codes? List them.
- A theatre logbook, departmental register, pathology database, pharmacy dispensing record?
- A free-text search of the electronic record? Give the exact search terms.
Then the question almost nobody answers: what would this method have missed? Coding-based identification misses miscoded and uncoded cases, and coding accuracy varies systematically — by complexity, by clinician, and often by whether a diagnosis affects reimbursement. A logbook misses anything not entered.
If you validated your identification method against another source, report the overlap. That single check separates a careful review from a convenient one.
4. State inclusion and exclusion criteria, with numbers
Criteria alone are not enough — give the attrition. How many records were identified, how many excluded, and for what reason at each step.
The search identified 214 records. Of these, 38 were excluded for a final diagnosis other than the study condition, 21 for age under 18, and 12 for follow-up shorter than 90 days, leaving 143 for analysis.
A flow diagram is better still, and increasingly expected.
5. Describe the abstraction form
Variables should be defined before abstraction begins, on a structured form. Say that this is what you did, list the variables, and — critically — give the operational definition of anything that is not self-evident.
"Post-operative complication" means nothing until you say whether it is Clavien-Dindo graded, over what window, and whether readmission counts. "Hypertension" needs a definition: recorded diagnosis, current antihypertensive, or a threshold reading. Different definitions produce different prevalences from identical notes.
If you used a standard instrument or classification, cite it.
6. Say who abstracted, and how they were trained
How many abstractors, what their clinical background was, and what training or calibration they received before starting. A single junior abstractor working from an undefined form is a different study from two trained clinicians using a piloted instrument, and readers deserve to know which they are reading.
Piloting the form on a small sample first is standard practice and worth reporting: it is where ambiguous definitions surface.
7. Report blinding — or its absence
Were abstractors blinded to the study hypothesis, to exposure, or to outcome status?
Complete blinding is often impossible in a chart review, and saying so is fine. What is not fine is silence, because unblinded abstraction is a real source of bias: a reader who knows the hypothesis interprets ambiguous notes in its direction, usually without noticing.
Partial measures are worth reporting — abstracting exposure and outcome variables in separate passes, or having a second person abstract outcomes without access to exposure status.
8. Report inter-rater reliability
Where two or more people abstracted, report agreement — Cohen's kappa for categorical variables, or intraclass correlation for continuous ones — and say how disagreements were resolved. Adjudication by a third reviewer is the usual mechanism; consensus discussion is also acceptable if stated.
Where one person abstracted everything, the honest route is a re-abstraction check: a second person independently repeats ten to twenty per cent of records, and you report agreement on the key variables. This costs a few hours and substantially strengthens the paper.
9. Handle missing data explicitly
Missing data is the defining feature of retrospective work, not an incidental problem. Nobody documented with your variables in mind.
Report, for each key variable, how much was missing. Then state how you handled it: complete-case analysis, imputation with the method named, or a category for "not recorded".
And distinguish carefully between not documented and not present. A note that does not mention nausea does not establish that the patient had none. Treating absence of documentation as absence of the finding is the most common analytical error in chart reviews, and it systematically biases prevalence downwards.
10. State ethics approval and consent status
Give the approving committee and reference number, and state whether individual consent was obtained or formally waived. A waiver of consent is standard for retrospective record review, but it must be granted, not assumed.
Also state your data protection arrangements: when identifiers were removed, whether analysis used a de-identified dataset, and where data were held.
Use STROBE — and RECORD where it applies
A chart review is an observational study, so STROBE applies. Where your data come from routinely collected health data or administrative databases, use RECORD, the STROBE extension built for exactly that situation. RECORD adds items on database provenance, the codes and algorithms used, linkage between datasets, and any validation of those codes.
Say which you followed in the Methods, and submit the completed checklist.
Limitations that must appear
Reviewers will raise these. Raising them first is better:
- Data were recorded for clinical care, not research, and were not standardised.
- Documentation quality varies between clinicians and over time.
- Undocumented does not mean absent.
- Case identification depends on coding or record-keeping accuracy; some eligible patients were probably missed.
- Unmeasured confounding is likely — you can only adjust for what was written down.
- Single or few centres limit generalisability.
- Where abstraction was unblinded, interpretation bias is possible.
The underlying point
A chart review cannot be made into a trial by better writing. What good reporting does is let a reader see exactly what was done, judge the biases for themselves, and calibrate their confidence accordingly.
That is achievable, and it is the difference between a paper reviewers trust and one they cannot evaluate. The work is mostly done before abstraction begins — a defined form, agreed definitions, a documented search — which is the strongest argument for planning the Methods section before collecting the data rather than after.
See also choosing between a case report, case series and cohort study and our guide to reporting guidelines.