Why NHS hospital co-morbidity research may be wrong: how clinical coding fails to identify the impact of diabetes mellitus on cancer survival

Abstract

Background: Significant volumes of research rely on secondary care diagnostic coding to identify comorbidities however little is known about its accuracy at a population level or if this influences subsequent analysis.

Methods: Retrospective observational study utilising real world data for all cancers, prostate cancer and breast cancer patients diagnosed at Leeds Cancer Centre from 2005 and 2018. Three different data definitions were used to identify patients with diabetes in each cohort: (1) clinical coding alone, (2) HbA1c blood test alone (3) either clinical coding or abnormal HbA1c. Cohort characteristics, diagnosis dates and Cox derived survival was compared across diabetes definitions.

Results: 123,841 cancer patients were identified including 13,964 with diabetes. Clinical coding failed to identify 14.6% of diabetic cancer patients with a temporal misclassification rate of 17.5%. Sole reliance on clinical coding overestimated the negative effect of DM on median survival across all cancers and 3.17 years in breast cancer.

Discussion: Clinical coding provides inaccurate diabetes diagnosis date and detection resulting in meaningful differences in analytic outcomes. This supports the use of more detailed comorbidity data definitions. Results casts doubt over research reliant on hospital clinical coding alone and the generalisability of some comorbidity and frailty scoring systems.

Fig. 1. Graphical representation of the diabetic subgroups.

Note that the shape area of each subgroup is not scaled to the true numbers in the dataset.

Fig. 2. Overlap of identification by HbA1c and clinical coding.

Proportion of the diabetic cancer population that are identified by Clinical coding but not HbA1c (Uniquely Identified by Clinical Coding), HbA1c and not clinical coding (Uniquely Identified by HbA1c), or identified by both HbA1c and Clinical Coding (Universally Identified). The results are shown for all cancer patients and the population of cancer patients in the LTHT blood catchment area.

Fig. 3. Difference in first diabetic diagnosis flag from clinical records comparing clinical coding and abnormal HbA1c.

 = patients defined by both an abnormal HbA1c and diabetic clinical coding.  = patients defined by clinical coding as post-cancer diabetics but with abnormal HbA1c pre-cancer. Left of  = abnormal HbA1c earlier than clinical coding. Right of  = clinical coding earlier than abnormal HbA1c.

Fig. 4. Survival trajectories for each diabetic data definition and the overall survival for all patients in the all cancer cohort.

Survival curves have been plotted for cox models generated using only all patients, coded diabetic patients, abnormal HbA1c patients and diabetic patients identified by the hybrid method. In all cases estimates were adjusted for potential age, sex and deprivation confounding.

Fig. 5. Cox derived hazard ratios for the impact of diabetes.

Comparison between the estimated hazard across all three cohorts using each of the diabetes definitions. Horizontal dashed line indicates a hazard ratio of 1.0, which is the threshold at which the estimated hazard is equal with and without diabetes