Internal and External validation of the IDIOM score for predicting the risk of gastro-intestinal malignancy in iron deficiency anaemia
Authors: Almilaji, O., Webb, G., Chapman, T.P., Williams, E.J., Shine, B.S.F., Ellis, A.J., Docherty, S. and Snook, J.
Conference: NCRI Cancer Conference
Dates: 2-3 November 2020
Abstract:Background Gastrointestinal (GI) malignancy is a common finding in iron deficiency anaemia (IDA), with a prevalence of about 8%. Using two large datasets from Dorset, we have previously reported and internally validated a model for predicting the risk of GI malignancy in IDA – the IDIOM score. This is based on four independent and objective clinical parameters - age, sex, mean corpuscular volume (MCV), and haemoglobin concentration (Hb). This study aims to assess the performance of the predictive model applied to an unrelated external validation dataset.
Method The external validation dataset was derived from a different population (in Oxford), collected under different circumstances (from fast-track referrals), and comprised a total of 1118 patients with confirmed IDA. The data were anonymised prior to analysis. The logistic regression model based on the training data was used to predict the GI malignancy risk in this new dataset. Due to the imbalance between the “positive” and “negative” GI malignancy numbers, geometric mean (G mean), and negative predictive value were used to assess the performance of the model.
Results The characteristics of the external validation dataset differed from those of the training dataset, with lower mean Hb in particular. Using the regression model to calculate predicted GI malignancy risk, a threshold risk of 7.43% maximised the G mean in the training dataset (69%) and gave a comparable value in the external validation dataset (61%). At this threshold, sensitivity and specificity were 76% and 63% respectively in the training dataset and 84% and 44% in the external validation dataset. A predicted risk threshold of 1.5% was the largest value to give a negative predictive value of 100% in the training dataset and gave an identical result in the external validation dataset.
Conclusion This external validation exercise has demonstrated that the model underlying the IDIOM score is robust in predicting the risk of underlying GI malignancy in a large IDA dataset collected in a different clinical setting.
Impact statement Ultimately, validating the model would help to using it to rationalise the use of investigational resources in IDA, by fast-tracking high-risk patients and, with appropriate safeguards, avoiding invasive investigation altogether in those at ultra-low predicted risk.
Source: Manual