Leveraging free-text diagnoses to identify patients with diabetes mellitus, obesity or dyslipidaemia - a cross-sectional study in a large Swiss primary care database.

Background: Electronic medical records (EMRs) in general practice provide various methods for identifying patients with specific diagnoses. While several studies have focused on case identification via structured EMR components, diagnoses in general practice are frequently documented as unstructured free-text entries, making their use for research challenging. Furthermore, diagnoses may remain undocumented even when evidence of the underlying disease exists within structured EMR data.

Objective: This study aimed to quantify the extent to which free-text diagnoses contribute to identifying additional cases of diabetes mellitus, obesity and dyslipidaemia (target diseases) and assess the cases missed when relying exclusively on free-text entries.

Methods: This cross-sectional study utilised EMR data from all consultations up to 2019 for 6,000 patients across 10 general practices in Switzerland. Diagnoses documented in a free-text entry field for diagnoses were manually coded for target diseases. Cases were defined as patients with a corresponding coded free-text diagnosis or meeting predefined criteria in structured EMR components (medication data or clinical and laboratory parameters). For each target disease, prevalence was calculated along with the proportion of cases identified exclusively via free-text diagnoses and the proportion missed when using free-text diagnoses alone.

Results: The prevalence estimates for diabetes mellitus, obesity and dyslipidaemia were 8.8%, 16.2% and 38.9%, respectively. Few cases relied exclusively on free-text diagnoses for identification, but a substantial proportion of cases were missed when relying solely on free-text diagnoses, particularly for obesity (19.5% exclusively identified; 50.7% missed) and dyslipidaemia (8.7% exclusively identified; 53.3% missed).

Conclusion: Free-text diagnoses were of limited utility for case identification of diabetes mellitus, obesity or dyslipidaemia, suggesting that manual coding of free-text diagnoses may not always be justified. Relying solely on free-text diagnoses for case identification is not recommended, as substantial proportions of cases may remain undetected, leading to biased prevalence estimates.