Unwarranted Variations in Healthcare: The Role of Business Intelligence

Resumen

Healthcare is one of the most impactful sectors, affecting strongly all aspects of the society ranging from medical services to social, governmental, business, and economic implications. Currently, there is an emerging concern of chronic diseases, the ageing and higher mobility of care professionals, which accentuates the need for national policies to enhance the availability, approachability, eligibility and quality of the care system. These changes are driving the cost of healthcare to unsustainable levels. On the other hand, it is well known that healthcare systems suffer from service delivery variations known as unwarranted variations. These are the variations that are not explained by patients’ needs or preferences. Existence of such variations indicates that the best care has not been delivered or that resources have not been appropriately deployed in healthcare organizations. These variations can occur in terms of treatment quality, expenditure, supply of health resources, etc. Detecting and reducing these variations is one of the main paths to decrease the healthcare expenditures without compromising the care quality. The correct identification of the unwarranted variations is the corner stone to detect the underlying factors and designing effective interventions to reduce the disparities. The proper detection of the unwarranted variations has two important antecedents: first, timely access to quality data and second, reliable risk adjustments. As not all variations are unwarranted, risk adjustment methods are used to separate the “good” variations from the total observed variations. These aspects are even more crucial when moving from larger to smaller levels of analyses. We believe that Business Intelligence (BI) is a prominent remedy for both of these vital elements. BI as a system that encompasses data collection, storage, analyses and reporting, has significant potentials in providing timely and quality data for unwarranted variations research. Moreover, BI can incorporate novel advanced analytics to better separate unwarranted variations from the total observed variations. However, literature reports that the success rate of BI in healthcare is significantly lower compared to other sectors. In order to cope with the above challenges, five studies categorized in three phases are developed in this PhD thesis. In the first phase, the cross-regional variations in healthcare performance in Spanish provinces are assessed, examining patients’ average length of hospital stay, mortality rates, readmission rates and hospital-acquired infection rates. The crude rates for the indicators have been obtained from the administrative data acquired by Spanish public hospitals, for non-surgical patients diagnosed with diseases in six Major Diagnostic Categories. These rates have been adjusted for gender, age, type of admission, severity of the condition and case-mix weights. The adjusted rates have been used as the input for a two-step clustering process, splitting the provinces into three groups based upon their performance. The results demonstrate that the clusters perform differently for mortality, average length of hospital stay and hospital-acquired infections rates, while they perform similarly for readmission rates. Moreover, the performance pattern by type of diagnostic category is similar for all the three clusters. While evidence in the literature regarding the adverse effects of length of stay reduction on readmission and mortality rates are mixed and inconclusive, our findings demonstrate that many regions in Spain can shorten the length of stay without risk of such advert effects. Similar performances across the diagnostic categories suggest that influencing factors are common and healthcare decision-makers should pay more attention to indicator-specific interventions rather than disease-specific ones. Next, we focused on the factors influencing potentially avoidable hospitalizations. The adjusted hospitalization rates were assessed for three common ambulatory care sensitive conditions -known to be responsible for potentially avoidable admissions-. The adjusted rates of the 17 Spanish autonomous communities were collected for the period between 2007 to 2015. Fixed effect model is fitted with the data that mitigate the bias caused by time-invariant omitted or unobserved factors. The findings show that consultation with general practitioners per capita is associated with less hospitalization rates for all three conditions. Moreover, specialists’ consultations per capita and the proportion of people visiting specialists is associated with reduced hospitalizations in two of the conditions. This is an indication that primary care cannot solely be held accountable for all preventable hospitalizations. In the second phase, we developed a comprehensive maturity model that suggest the priorities to be followed for progressively developing a BI solution. Through collaborative research in several Italian hospitals, we demonstrated that the successful implementation of BI is an evolutionary and complex path. By developing a maturity model specific for BI in healthcare, we provide managers with means that guide this evolutionary implementation. We identified a comprehensive set of BI dimensions and indicators that should be considered during the implementations. Finally, in the third phase, we focused on advanced analytics that can be embedded in the BI systems and support the investigations in unwarranted variations. We first developed a novel Association Rules Mining (ARM) algorithm named Length-Sort that can discover the maximum-length association rules more efficiently. The efficiency improvement enables the analysis of larger health datasets. Next, we demonstrate the application of Length-Sort on England Quality and Outcome Framework (QOF) dataset which contains general practices’ performances in terms of various evidence-based indicators. Our results show that ARM and in particular, Length-Sort can be an explorative alternative to the traditional risk adjustment method, which can be used in health policy and unwarranted variation research as a hypothesis generation tool to detect performance patterns especially, in the more granular levels of analyses.