Identifying long-term conditions in New Zealand general practice using structured and unstructured data: a cross-sectional study.

Abstract

Objectives: This study examined whether incorporating free-text entries into structured general practice records improves the detection of long-term conditions (LTCs) and multimorbidity (MM) in New Zealand (NZ) general practices.

Methods: Data from 374 071 deidentified individuals in general practices were analysed to identify 61 LTCs. Structured data were extracted using Read codes from a national master list, and clinical raters independently identified condition-related free-text, including synonyms, negation terms and common misspellings in randomised samples. Keywords were categorised and refined through ten iterative tests. Programmatic text classification was developed and assessed against gold-standard clinician ratings, using sensitivity, specificity, positive predictive value (PPV) and F₁-score.

Results: A quarter of general practitioner classifications contained either unrecognised Read codes or consisted of free-text only. Clinician inter-rater reliability was high (kappa ≥0.9). Compared with clinical gold standard, text classification yielded an average sensitivity of 88%, specificity of 99% and PPV of 95%, with an F₁-score range of 82%-95%. Incorporating free text increased LTC prevalence from 42.1% to 46.3%, reducing misclassification of MM diagnoses by identifying 12 626 additional patients with MM and 15 972 additional patients with at least one LTC.

Discussion: In the course of workflow, general practitioners face barriers to accurate LTC coding or may simply annotate with text-based descriptions. Programmatic text classification has demonstrated high performance and identified many more patients receiving LTC care.

Conclusions: Combining structured and unstructured data optimises MM detection in NZ general practices and has the potential to improve case management, follow-up care and allocation of healthcare resources.