Process and outcome of care analysis using administrative datasets

Abstract

The many established and mandatory data capture systems in healthcare offer the opportunity for' organisational learning. ' This thesis explores the use of such administrative data (AD), and specifically hospital data sets that have been collected primarily for funding and reporting, for analysis of process and outcome of care for strokes and transient ischaemic attacks (TIAs). As a first step the problems encountered when selecting stroke and TIA cohort from multi-year, state-level, hospital ADs were examined. The cohort selection relies on the recorded patient diagnoses information: International Classification of Disease (lCD) codes and/or Diagnosis Related Groups (DRGs). In the study years the difficulties in accurately selecting the cohort were found to be due to various interpretations of codes and DRGs, the placement of codes to groupers and the different classifications used. Discovered irregularities were reported to the appropriate government department. It became obvious that experienced coders' and clinicians' perspectives along with good knowledge of the changes to codes and groupers are required for accurate cohort selection. Since information about the number and types of non-principal diagnoses (NPDs) would be useful for predicting outcome and hence patient management, the NPDs were examined using the associated prefixes (that identify the onset or relevance of each NPD to the episode) recorded in three fiscal-years of Victorian admitted datasets. The study revealed that the presentation of TIA and strokes are varied and complex, with confirmation of some known relationships and other new ones revealed. The results showed that the prefix categories accompanying NPDs can help to better define the nature of the presentation and thus explain some ofthe observed outcomes. Apart from improving the definition and collection guidelines it is important for the regular statewide audits to also give feedback that helps improve the quality of prefixes. The treatment of patients with multiple conditions is a norm rather than an exception. Following the review of generic multi-morbidity instruments, it was concluded that most of the basic design features were acceptable. However important construct validity needed addressing, especially the embracing of a more precise and standardised definition such as the emerging condition/disease-specific complexity-of-patientpresentation construct, and a clearer delineation of application and limitations. Therefore a stroke-specific ICD-l O-based complexity-of-stroke-patient-presentation measure using AD was developed and tested. Condition-groups with a potential to contribute to the complexity-of-stroke-patient-presentation were identified and the mapping codes selected. Based on their relative potential to contribute towards the complexity-of-strokepatient- presentation, an index value was assigned to each condition-group. The index values of non-principal diagnoses were used to calculate the Total Complexity Weight for each patient. Convergent validity was tested using a derived 'pseudo gold standard.' U sing a state AD the alignment of the complexity score to patient factors and outcomes was established. Finally, all of the above methodological improvements were used in analysing the ADs. Since Bayesian Networks (BNs) provide an easily understood representation of causal relationships and support ad hoc exploration of impacts on local distributions, it was the method of choice. Two BN studies are reported. First, formal transfers in and out ofEDs with and without Stroke Care Units (SCUs) were analyzed. This revealed that teaching hospitals with SCU s, while achieving shorter length of stay, in fact deal with younger patients with lower overall patient complexity than non-SCU teaching hospitals. Second, ED triaging of suspected stroke patients who subsequently experienced an inpatient admission was analyzed. It was notable that 45% of TIAs were categorised as only 'Semi-urgent,' indicating an opportunity to improve emergency assessment ofTIAs. These studies demonstrated that the learning algorithm used with the hybrid BN when applied to ADs can reveal high-level details of the care journey and outcome. ADs are often the best available operational and historical data that are readily obtainable and relatively inexpensive. This project demonstrated that value can be drawn from these datasets to provide high-level insight into the process and outcome of care of a specific cohort of patient. In an era of diminishing resources better utilisation of these datasets should be encouraged. As electronic information systems are increasingly embraced, these collections need to be managed as valuable assets and powerful operational and patient management tools.