Statistical methods for classification in cardiovascular healthcare

Since a strong attention has been devoted in the last decades to healthcare management for social, medical and economical reasons, over recent years also the analysis, development and improvement of suitable tools for measuring the quality of care has become a research field of extreme importance. Within this context, since patient outcomes enable researchers (at least in part) to assess the quality of care, there has been a widespread diffusion of techniques for monitoring and evaluating the underlying processes generating such outcomes. This task is nowadays unavoidable in order to get a sensible improvement of healthcare services quality, as well as to contain economical costs. In order to achieve a suitable strategy for assessing healthcare performances, it is necessary to identify suitable operating protocols, as well as the functional competence of institutions, and then to monitor them over time. In this way, a continuous collection of data proved to be mandatory. With this respect, a dataset (clinical registry and/or administrative database) describes the process underlying itself. The more complex, structured and detailed is the dataset, the more difficult is the analysis required for its exploitation. In other words, monitoring healthcare systems through data collections asks for careful design of experiment and high quality of collected data, shared standards of collection and strict control on filling compliance and reliability of the data. Moreover, it calls for suitable statistical methods for analysis, modelling and predictions. Only in this way it is possible to derive models which are capable of realistic interpretations of the process underlying the dataset. Statistical techniques, when applied to measurement data, can be used firstly to highlight areas that would benefit from further investigation, then to model processes relating patterns of care, patients case-mix, hospital influences and outcomes of interest, and finally to make predictions on them. Statistics enables the researchers to identify variation within the process under observation. Understanding, modelling and then quantifying this variation are the first steps towards quality improvement. An important goal of Regione Lombardia (healthcare division) is the use of performance measures for monitoring cardiological and cardiovascular healthcare offer, as well as to assess institutions within the regional healthcare service in order to provide evidence for initiatives aimed at enhancing professional accountability in the public sector. Specifically, a Strategic Program, named “Sviluppo di nuove strategie conoscitive, diagnostiche, terapeutiche e organizzative in pazienti con sindromi coronariche acute” (www.salute.gov.it/ricercaSanitaria), has started in 2008 with, among others, the goals of (i) pointing out a comprehensive clinical and epidemiological picture of how Acute Myocardial Infarction (AMI) is treated in Regione Lombardia; (ii) assessing the effectiveness of AMI patterns of care, in order to invest in innovations starting from real epidemiological evidence and needs; (iii) exploiting administrative databanks for addressing clinical and epidemiological enquires; (iv) highlighting critical situations in healthcare delivery and then to improve hospital performances; (v) providing people in charge with healthcare government with decisional support based on statistical evidence and real time data. In order to address these issues, suitable methods to collect, analyze and model data are needed. The results of statistical analyses carried out on data arising both from clinical registries and administrative databanks may influence funding and policy decisions, and are used to generate feedback for providers. The providers’ profiling based on current data collections is a new way for improving quality of healthcare offer. This requires shared information-technology systems of data collection and advanced statistical methods able to classify providers, to quantify their effect on outcomes of interest at patients’ level, to analyse complex data arising from biomedical context and to make reliable predictions. Statistics is then of paramount importance in more than one step of the cardiovascular healthcare process, especially in supporting this new concept of “real-time” epidemiology based on observational clinical registries and administrative databanks. In fact, as shown and explained in the thesis, the statistician plays a central role during the design of experiment, carries out the monitoring of data collection, evaluates the process and produces a feedback for involved players, elaborates models necessary for providers’ profiling, classification and outcomes prediction. The decisional support provided by statisticians is evidence based and it is based on real epidemiological evidence and needs, involving low cost data sources, i.e., real-time and sustainable from economic perspective. In this thesis a general approach to model fitting aimed at clustering is considered, focusing on grouped (longitudinal) data arising from healthcare context, where examples of grouping factors are hospitals (and diseases) with respect to patients or patients themselves with respect to their own measurements over time. The main goal of the thesis is then to present a number of statistical techniques for the analysis of such data, in order to provide methods for supporting decisions of people in charge with healthcare government. Clinically speaking, we will focus specifically on problems related to the improvement and optimization of pattern of care for patients affected by Acute Coronary Syndromes. On the other hand, the main statistical topic we will deal with is clustering carried out starting from random effects models estimation. In fact, mixed effects models are used in a wide variety of biostatistical contexts, and can be analysed both from a classical and Bayesian viewpoint. We can distinguish two types of applications: those in which the random effects are nuisance parameters, and are not of direct interest and those in which the individual effects are of paramount interest. Although mixed effects models are used in many applications in medical statistics, for our scopes the most interesting and new application is the problem of hospital comparisons using routine performance data. Among other benefits this approach provides a diagnostic criterion to detect clusters of providers with unusual results.

Scopo di questo lavoro di tesi è sviluppare un approccio metodologico generale allo studio di modelli per la classificazione di dati raggruppati, derivanti dall’ambito clinico sanitario in generale. Esempi di raggruppamento all’interno di tale contesto sono dati dall’ospedale di ricovero o dalla patologia, qualora le unità statistiche di interesse siano i pazienti; in alternativa, i pazienti stessi possono essere considerati un fattore di raggruppamento per misure ripetute dei propri outcome clinici nel tempo. Obiettivo principale del lavoro è dunque presentare diverse metodologie statistiche vagliate e sviluppate per l’analisi di tali dati, allo scopo di fornire un supporto alle decisioni a chi si occupa di politica e pianificazione sanitaria in Regione Lombardia. Da un punto di vista clinico, l’attenzione si concentra su problemi relativi al miglioramento nonché all’ottimizzazione dei percorsi terapeutici a cui sottoporre i pazienti affetti da Sindromi Coronariche Acute. Da un punto di vista statistico, quanto appena detto si concretizza nello studio di tecniche per la classificazione delle prestazioni sanitarie a partire dalle stime degli effetti fissi e casuali presenti nei modelli di volta in volta considerati. I modelli a effetti misti, infatti, sono ormai utilizzati in un gran numero di problemi di natura clinica e biologica. All’interno di questo lavoro di tesi, ne viene analizzato e considerato sia l’approccio classico che quello Bayesiano. In particolare, si focalizza l’interesse proprio sullo studio di come gli effetti casuali permettano di cogliere e quantificare l’effetto del fattore di raggruppamento sulla risposta. Nello specifico del contesto clinico in cui si inserisce questo lavoro di tesi, siamo interessati ad effettuare una valutazione comparativa del comportamento degli ospedali nonché un confronto della loro efficienza ed efficacia terapeutica nella gestione dei pazienti affetti da Sindromi Coronariche Acute, a partire dall’evidenza proveniente dalle raccolte dati cliniche e amministrative. Oltre ai comprovati benefici che un monitoraggio continuativo dell’operato delle strutture può portare, l’utilizzo di tecniche statistiche avanzate nella modellizzazione della dinamica processo-outcome clinico può fornisce criteri decisionali per l’individuazione in tempo reale di comportamenti anomali tra le strutture erogatrici di servizi, permettendo di localizzare aree e situazioni che necessitano interventi atti al miglioramento della qualità del servizio e del processo di cura.