Re-imagining the surveillance of invasive mould diseases

Abstract

Invasive mould diseases (IMDs) have significant health and economic costs for immunocompromised patients. IMDs are now the predominant fungal pathogens in haematology-oncology patients who also carry the greatest burden of fungal infections overall. This thesis captures the high hospitalisation cost of invasive fungal diseases (IFDs) based on detailed patient level data. It explores the optimal methodology to cost these infections and postulates a novel resource metric in costly antifungal drug consumption as an alternative to currency estimates. It examines the comparative effectiveness of antifungal prophylaxis in haematology-oncology patients at high risk for IMDs documenting over a decade, the declining incidence of IMD following the adoption of mould active antifungal prophylaxis with reductions in other clinically relevant outcomes such as empiric antifungal therapy and non- specific pulmonary infiltrates also observed. Our analyses of azole and liposomal amphotericin prophylaxis go beyond demonstrating the evolving epidemiology of fungal infections in response to changing therapeutic practices but also illustrate that knowledge of local epidemiology informs clinical decision making. The latter, culminating in clinical practice recommendations regarding the optimisation of posaconazole for either prophylaxis or treatment of established IFDs and, in a broader context, recommendations strengthening antifungal drug stewardship programs in hospitals.

Despite the health and cost implications of IMDs and effort invested in preventing them, prospective continuous surveillance is not routinely performed in many hospitals. Reasons for this omission include cost and the absence of an easily identifiable laboratory prompt. Traditional approaches to surveillance, which are reliant on bedside review, administrative or laboratory- based methods, are resource-intensive activities, error prone and subject to either under-reporting and/or variability in case ascertainment.

This thesis will argue the case for prospective surveillance of IMDs and provide a potential technological solution to facilitate its practice in hospitals. From a surveillance standpoint, the primary screening method is critically important in order to maximise case finding while minimising its cost and effort. For IMDs however, the optimal screening method is undefined. This thesis considers targeting computed tomography (CT) reports as an appropriate screening method for IMD surveillance. CT is a key diagnostic modality for IMDs stipulated in consensus guidelines; pulmonary involvement is present in the overwhelming majority of cases; it is widely available in hospitals and being a non-invasive test it is uniformly performed when IMDs are suspected with results available in a timely fashion. Non-culture based tests (NCBTs) such as galactomannan (GM) or polymerase chain reaction (PCR) are less suitable as screening tools due to their variable availability, delays in turn- around and a diminished sensitivity in the presence of concomitant mould- active antifungal agents.

This thesis describes the development of a text classifier that uses natural language processing (NLP) for the first time to flag CT reports supportive of IMDs. NLP is a computational method for analysing human language that has been applied for the detection of a variety of medical conditions, but not IMDs. As a high-throughput technology, it has the potential to identify CT reports with suspected IMD in real-time. Thus, it may deliver to hospitals a feasible, sustainable and cost-effective solution to IMD surveillance with minimal interruption to routine clinical workflow.

Text analytical tools are a means of unlocking the wealth of patient-level data that is largely confined to unstructured (i.e., free-text) documents in health care. Modernisation of data management and an investment in data infrastructure could help the health industry keep pace with changing clinical practice while also supporting large scale comparative effectiveness studies of ‘real-world’ patients who are free from the protocol-driven biases of clinical trials. In future, the partnering of enormous volumes of routinely collected clinical data with genomic or molecular advances could help progress bioinformatics which predict disease and personalise treatments that are both effective and cost-beneficial.