Background
Over the last two decades, there has been increasing interest in the development of performance indicators in the attempt to promote accountability in health services [
1]-[
3]. These measures may concern different aspects of the system and reflect different objectives. ‘Process’ measures, as surrogate outcome indicators, have been used to assess whether specific care processes recommended in clinical guidelines are administered, such as intervening within 48 hours of a hip fracture (HF). ‘Outcome’ measures, such as the 30-day mortality rate after hospital admission for acute myocardial infarction (AMI), have been used to evaluate the effectiveness of health care processes.
In this respect, hospital discharge records have been an essential source of information for comparing health outcomes among hospitals [
4]. They are widely available and represent a cost-effective source of information for monitoring health care quality in clinical practice over large populations and across a wide variety of conditions and procedures [
5],[
6]. However, the amount of patient-level information collected in these archives is limited, generally consisting of age, gender, discharge diagnoses and main procedures. These data do not allow observers to characterise the acute clinical severity of the patient, but only to identify specific diseases as ‘comorbidities’ (e.g., chronic pre-existing conditions that increase the a priori risk that the subject will incur adverse short-term health outcomes) [
7],[
8]. For these reasons, the ability of administrative data to adjust for the severity of illness and to provide unbiased estimates of expected mortality rates at the hospital level has been criticised [
9].
On the opposite extreme, clinical or laboratory data abstracted from medical records, when available, represent a good alternative because they may better account for the pre-hospital or pre-operative severity of illness, they may distinguish comorbidities (conditions already present at the time of admission or the procedure) from complications (conditions arising during hospitalisation or during the procedure) and they do not limit the number of reported diagnoses, so they avoid differential reporting of conditions according to the baseline severity of the patient [
10],[
11]. The main drawbacks of the clinical archives are that their reliability may differ between hospitals and that it is difficult to obtain a large amount of clinical data at an affordable cost. Therefore, it is difficult to implement risk-adjustment methods based on these data in a systematic way [
5],[
10].
Many authors and physicians advocate integrating both types of data to take full advantage of their relative strengths [
12]. However, this strategy is often not feasible in clinical practice because of the difficulty obtaining timely and complete information on acute clinical severity. Thus, some investigators [
13] have questioned whether it is possible to identify a limited number of laboratory or clinical data points that would be affordable and easy to collect from electronic medical archives and could be used to improve risk adjustment of inpatient mortality for different clinical conditions or procedures. In other words, as Johnston and colleagues asked, ‘Is there a low-cost way to improve the risk adjustment of administrative data?’ [
5].
Following these principles, medical professionals from different clinical areas and public health authorities began an audit activity in 2006 in the Lazio Region, Italy, with the aim of complementing the Hospital Information System (HIS) with a few selected clinical variables chosen to better characterise the acute severity of patients admitted for specific conditions. Ultimately, Coronary Artery Bypass Grafting (CABG), AMI and HF were selected, and a few clinical or laboratory data points were identified for each of them. The collection and transmission of this information became mandatory for all public and private hospitals in the Lazio region (where Rome is located) in 2008, and the new data became part of the new HIS.
Another information system with strong potential for comparative effectiveness research is the Regional Drug-Dispensing Registry (PHARM), which includes individual records for each drug prescription dispensed in public and private pharmacies. Some studies have evaluated the use of pharmacy dispensing data for predicting healthcare utilisation [
14] or to identify patients with chronic conditions [
15] but not specifically to control for confounding.
However, there is extensive scientific literature on the performance of ‘a priori’ comorbidity scores based on outpatient pharmacy dispensing data, such as the Chronic Disease Scores [
16]-[
18]. Despite their popularity, these indices have limited utility in controlling for confounding because, like all summary scores, they assume a fixed relationship between comorbidities and the outcome, even though this relationship is likely to differ between populations. In fact, the risk-adjustment process should involve the construction of empirical illness severity and comorbidity measures specific to the study population. From this perspective, the scientific literature currently has significant gaps in terms of the evaluation of different types of empirical models for risk-adjustment procedures.
The integration of diagnosis-based models with medication-based predictive models is expected to result in greater predictive power and more exhaustive control of confounding, by modelling the complex relationships between diagnosed comorbidities and the presence of any pharmacological therapy, taking into account its benefit and harms. Extending these approaches to different diseases and conditions for developing and applying risk-adjustment models can provide new evidence to evaluate whether data on hospital performance are credible or methodologically flawed.
In the present study, we analysed two indicators, 30-day mortality after AMI admission and surgery within 48 hours after HF admission, with two objectives: 1) to compare a risk-adjustment model based on hospital discharge data only with a model including clinical variables and drug prescription information and 2) to investigate whether the two risk-adjustment procedures lead to different conclusions about hospital comparisons.
Discussion
The present study was designed to quantify the additional contribution of clinical variables and drug prescription information in predicting short-term outcome rates and profiling hospitals compared to the use of hospital discharge data alone. We found that the risk adjustment improved considerably for the AMI cohort after the three sources of data were integrated, whereas the two risk-adjustment models performed equally poorly for the HF cohort. However, hospital profiling was not affected by the use of clinical variables and drug prescriptions for the HF cohort, whereas for the AMI cohort, some differences in hospital profiles were observed even if the ‘low performing’ and the ‘best performing’ institutions were the same regardless of the risk adjustment model applied.
The optimal approach to producing hospital outcome reports relies on collecting valid information to provide an accurate risk adjustment. This approach requires medical chart abstraction, which is expensive and therefore has not been widely implemented by public reporting agencies. Using administrative data for public outcomes reporting offers several advantages, including minimal data collection costs and the ability to produce reports for a large number of procedures and conditions [
26]. However, these data do not capture important clinical information about the acute severity of the patient and do not distinguish between the conditions that were present at admission and the complications that occurred during hospitalisation [
7],[
27],[
28]. Many authors have advocated for identifying a limited number of affordable and easily accessible laboratory or clinical data points from electronic medical archives that would improve risk-adjustment models of inpatient mortality for different clinical conditions or procedures [
12]. On this basis, in the present study, we used clinical information from the upgraded version of the HDR, including a few selected clinical variables chosen to better characterise the acute severity of patients admitted for AMI and HF. The Lazio HIS information is widely available and high in quality, and it represents a highly cost-effective solution for monitoring health care quality in clinical practice over large populations and across a wide variety of conditions and procedures.
The conditions to integrate from the discharge abstracts and the clinical variables to add were derived from an extensive audit activity that was conducted beginning in 2006 in the Lazio Region, Italy by medical professionals from different clinical areas and public health authorities. They opted for AMI and HF, conditions that pose significant public health problems, and identified a few clinical parameters (blood pressure and INR) that could be detected at affordable costs and are considered valid and reliable markers of acute severity.
The present study has some important strengths. It is the first study conducted in Italy with the specified aim of comparing the performance of risk-adjustment models with and without clinical variables and drugs prescription information in the context of hospital profiling and comparative outcomes research in general. The high number of patients investigated, the accuracy in the selection of the cohorts and the study outcomes, the consolidated statistical strategy, and the replication of similar findings for different clinical conditions are important elements of internal and external validity.
A limitation of this study is the generalizability of our results due to selection of only hospitalizations from Lazio region, however the large number of residents and hospitals in this region minimize the variability of case-mix between the admission in Lazio hospitals and in other Italian hospitals. Other limitations should also be acknowledged, especially the marked variability in the coding accuracy of current health care information systems. This issue is critical for ensuring accurate risk adjustment and thus reliable comparative quality ratings [
29]. However, data derived from health information systems are currently utilised to compare inpatient care outcomes in Italy [
30],[
31] and have proved to be an accurate source for healthcare research and a reliable data source for adjusting for risk factors [
8],[
32].
Moreover, as in the case of HF, risk-adjustment models may not be able to predict the study outcome (surgical treatment), even when they include valid clinical information on the severity of disease and drug prescriptions. This result simply means that the determinants of the outcome should be sought among the characteristics of the hospital and health care, which is the eventual purpose of hospital comparisons.
More generally, in outcome comparisons between hospitals, where each hospital represents a level of exposure, potential clinical confounders cannot produce important changes in the adjusted measures of association if these factors are not heterogeneously distributed between hospitals, even when they are strongly associated with the outcome under study.
Conclusions
In conclusion, the present study represents the first effort in Italy to compare the performance of risk-adjustment procedures incorporating clinical variables and drug prescriptions in predicting short-term outcomes and in profiling hospitals on the basis of the predicted outcomes. We found that the available clinical variables and drug prescription information were important complements to the hospital discharge data for characterising the acute severity of the patients for one of the two conditions we analysed. However, when the output of the predictive models was used to compare the hospitals on the basis of their risk-adjusted outcomes, the contribution of the clinical variables and drug prescriptions was always negligible. We hope that this approach will be replicated in other studies, for other clinical conditions, and for different clinical parameters and alternative analytical procedures to better interpret the present results and to better understand the trade-off between the costs and the advantages of including relevant clinical variables in systematic health information systems.
Data used for the study
The data used for the study are not openly available. The Department of Epidemiology has been authorised by the Regional Health Authority to use the data.
Authors’ contributions
PC: study conception and design, data analysis, interpretation of data. MDM: study conception and design, interpretation of data. DF: study design, data analysis, interpretation of data. PA: study conception and design, interpretation of data. MD: study conception and design, interpretation of data. CAP: study conception and design, interpretation of data. All authors read and approved the final manuscript.
Competing interests
The authors declare that they have no competing interests.