Defining and assessing the spectrum of tuberculosis (TB) disease: application to diagnosis and prognosis

We demonstrate not only that sensitivity and specificity by themselves offer a limited view into the performance of a test, but also that these metrics are not intrinsic to a test, but vary with disease prevalence.More generally, the contextual 'spectrum of disease' -which, in practice, mainly means the distribution of the times since infection in the population being tested.In exploring study design options, we note the large number of assumptions that are required to fully specify details of conventional 'power calculations', suggesting that this is not a clear cut approach to choosing sample sizes.Given data generated by a well-understood biological process, we find that formal criteria driven by automated methods for optimising analysis, such as the least absolute shrinkage and selection operator (LASSO), provides little or no advantage over intuitively chosen diagnostic threshold criteria.A detailed mapping of data fields, linking the DS-TB and DR-TB treatment databases is produced, supporting consistent analysis across both types of TB, and facilitating analysis of some of the rich but complex structures in the DR-TB database.Although a significant number of drug-resistant TB patients do not have a recorded treatment outcome, unfavourable treatment outcomes, such as death, are found to be alarmingly common, and significantly associated with HIV status, history of previous TB treatment, age, and resistance patterns.Better management of TB, a persistent and complex infectious disease, will require substantial additional research to be conducted in the coming years.It is hoped that this thesis will provide a meaningful resource to workers in this field, assisting them with numerous aspects of the search for better characterisation of the spectrum of TB; in particular, ways to access and analyse critical data and optimally design data gathering and analysis.