Doctoral Degrees (Chemical Pathology)

Permanent URI for this collection

https://scholar.sun.ac.za/handle/10019.1/3788

Browse

Browsing Doctoral Degrees (Chemical Pathology) by Subject "Diabetes -- Forecasting -- Mathematical models"

Now showing 1 - 1 of 1
  • Thumbnail Image
    Item
    Methodological issues around the validation of models for predicting diabetes risk in developing countries
    (Stellenbosch : Stellenbosh University, 2016-12) Masconi, Katya Laura; Kengne, Andre Pascal; Erasmus, Rajiv T.; Matsha, Tandi E.; Stellenbosch University. Faculty of Medicine and Health Sciences. Dept. of Pathology: Chemical Pathology.
    ENGLISH SUMMARY: Background: Multivariable diabetes risk prediction models have the potential to contribute to screening strategies, combining several risk factors to predict undiagnosed diabetes or future risk of developing diabetes. The focus of this study is the prediction of undiagnosed diabetes and diabetes risk prediction in a developing country where no population-specific diabetes risk prediction model currently exists. Existent models have been developed in unrelated populations with different disease prevalence, predictor weightings and methods used for risk factor determination and diabetes diagnosis. For accurate diabetes risk prediction in the mixed ancestry population of Bellville South, Cape Town, methodological issues regarding the validation and performance of these models needs to be addressed. Methodology: Cross-sectional data from the Cape Town Bellville South cohort was used for this study. Missing data in risk prediction research was investigated through a systematic review and a number of imputation methods were explored to deal with missing data in this dataset. Models were identified via recent systematic reviews and validated in the mixed-ancestry population. Discrimination was assessed and compared using the C-statistic and calibration was assessed via calibration plots. Model recalibration in diabetes risk prediction was investigated through a systematic review. In an effort to improve model performance in the new setting, model recalibration and updating strategies were used and performance was compared before and after implementation. Results: The study sample consisted of 1256 individuals, of whom 173 were excluded due to previously diagnosed diabetes. Of the final 1083 individuals, 329 (30.4%) had missing data. Deletion resulted in the lowest model performance and simple imputation, the simplest method, resulted in the highest model performance and was employed for further analysis. A systematic review highlighted the gross underreporting and mishandling of missing data in diabetes risk prediction research. Original model performance during validation was poor-to-average, with both over- and underestimation present: Cambridge [C-statistic: 0.67 (0.62-0.72); E/O: 1.81 (1.09-2.52)], Kuwaiti [C-statistic: 0.68 (0.63-0.73); E/O: 0.72 (0.43-1.12)], Omani [C-statistic: 0.66 (0.61-0.70); E/O: 1.28 (0.63-1.93)], Rotterdam [C-statistic: 0.64 (0.59-0.69); E/O: 0.54 (0.50-1.04)] and Simplified Finnish [C-statistic: 0.67 (0.62-0.71); E/O: 0.26 (0.13-0.39)] diabetes risk prediction models. Recalibration, as shown through a systematic review, was undertaken only in models predicting incident diabetes, and was reported in 22.9% of validation studies, with 77.8% achieving an increase in model performance. Updating results applied to this validation dataset showed an increase in both discrimination and calibration in varying levels across all five models. Overall, the re-estimation of the Cambridge diabetes risk model yielded the best model performance [C-statistic: 0.71 (0.67 – 0.75); E/O: 1.00 (0.86 – 1.17)]. Discussion and conclusion: The frequency of missing data, underreporting and mishandling of missing data, complexity of updating methods and overall model performance of validated models in new settings highlight the challenges in diabetes risk prediction research. This is the first validation study of prevalent diabetes risk prediction models in Sub-Saharan Africa and highlighted important methodological issues. While both simpler imputation and updating methods resulted in similar predictive utility when compared to more complex techniques, model performance was not increased sufficiently to suggest recommendation.