Medicalising normality? Using a simulated dataset to assess the performance of different diagnostic criteria of HIV-associated cognitive impairment

Abstract

Objective: The reported prevalence of cognitive impairment remains similar to that reported in the pre-antiretroviral therapy era. This may be partially artefactual due to the methods used to diagnose impairment. In this study, we evaluated the diagnostic performance of the HIV-associated neurocognitive disorder (Frascati criteria) and global deficit score (GDS) methods in comparison to a new, multivariate method of diagnosis.

Methods: Using a simulated 'normative' dataset informed by real-world cognitive data from the observational Pharmacokinetic and Clinical Observations in PeoPle Over fiftY (POPPY) cohort study, we evaluated the apparent prevalence of cognitive impairment using the Frascati and GDS definitions, as well as a novel multivariate method based on the Mahalanobis distance. We then quantified the diagnostic properties (including positive and negative predictive values and accuracy) of each method, using bootstrapping with 10,000 replicates, with a separate 'test' dataset to which a pre-defined proportion of 'impaired' individuals had been added.

Results: The simulated normative dataset demonstrated that up to ~26% of a normative control population would be diagnosed with cognitive impairment with the Frascati criteria and ~20% with the GDS. In contrast, the multivariate Mahalanobis distance method identified impairment in ~5%. Using the test dataset, diagnostic accuracy [95% confidence intervals] and positive predictive value (PPV) was best for the multivariate method vs. Frascati and GDS (accuracy: 92.8% [90.3-95.2%] vs. 76.1% [72.1-80.0%] and 80.6% [76.6-84.5%] respectively; PPV: 61.2% [48.3-72.2%] vs. 29.4% [22.2-36.8%] and 33.9% [25.6-42.3%] respectively). Increasing the a priori false positive rate for the multivariate Mahalanobis distance method from 5% to 15% resulted in an increase in sensitivity from 77.4% (64.5-89.4%) to 92.2% (83.3-100%) at a cost of specificity from 94.5% (92.8-95.2%) to 85.0% (81.2-88.5%).

Conclusion: Our simulations suggest that the commonly used diagnostic criteria of HIV-associated cognitive impairment label a significant proportion of a normative reference population as cognitively impaired, which will likely lead to a substantial over-estimate of the true proportion in a study population, due to their lower than expected specificity. These findings have important implications for clinical research regarding cognitive health in people living with HIV. More accurate methods of diagnosis should be implemented, with multivariate techniques offering a promising solution.

Conflict of interest statement

Competing Interests: The POPPY study is funded from investigator initiated grants from BMS, Gilead Sciences, Janssen, Merck and ViiV Healthcare to Imperial College London on behalf of Professors Alan Winston and Caroline Sabin. Alan Winston is a member of the PLOS ONE editorial board. This does not alter our adherence to PLOS ONE policies on sharing data and materials.

Figures

Fig 1. Convergence of the critical value to determine cognitive impairment as the sample size increases.

This model assumes a six test/domain model. The critical value from simulated data was the mean of 100 replicates.

Fig 2. Comparison of the cognitive domain correlation matrices for the HIV-positive and HIV-negative control groups from the POPPY study.

Visualisation of the inter-domain correlation matrices for the HIV-negative (panel a) and HIV-positive (panel b) participants of the POPPY study. Colour scale determined by Pearson’s r (scale to the right of the figure).

Fig 3. Histograms of a simulated study population with a 10% prevalence of cognitive impairment.

Panel a) 90% of the population are ‘normal’ and have a mean (standard deviation) T-score of 50 (10)–red. 10% of the population are impaired and have a mean T-score of 30 (10). Panels b-d: how the population is labelled by method used to define impairment.

Fig 4. Positive predictive value of each diagnostic method by prevalence of impairment in the simulated study population.