1. Introduction
The context and practice of clinical care have improved in recent years in terms of patient safety through advances in medicine and technology. However, the occurrence of adverse events (AEs), defined as an injury caused during the health care process, rather than by the underlying disease or condition of the patient, remains a significant issue [
1,
2,
3,
4]. The Institute of Medicine reported that medical errors cause more than a million injuries annually in the USA [
5], and that patient harm occurs in 25.1% admissions. Moreover, about half of these cases required intervention or caused permanent harm, and 63.1% were regarded as being preventable [
6].
Promoting patient safety usually involves a common strategy to detect patient safety events (PSEs), defined as an unexpected or unintended event that could have led or did lead to harm to the involved patient during the care process. Typical events might include patient falls; medication events such as prescription errors, dispensing errors, and drug administration errors; tube and line events such as inadvertent removal of an indwelling tube or line, etc. With this approach, studies have estimated an AE rate of 2.9–16.6% [
6], prompting the establishment of clear processes to detect, aggregate, and analyze PSEs to generate strategies to prevent the occurrence of PSEs, improve care quality, and promote a better patient safety climate [
7]. Typical endeavors by the healthcare institutions emphasize the establishment of an incident reporting system (IRS) and encourage reporting [
8,
9], while others have also encouraged analyzing PSEs to gain insights into systemic problems and preventive measures [
10,
11,
12].
Besides, the association between PSEs and healthcare utilization and costs has also been increasingly studied. A study in the period 2003–2004 reported the AE-related costs of over two billion Australian dollars annually [
13]. As there has been a trend of non-reimbursement for AE-derived medical expenses [
14,
15,
16], cost estimation and analysis of its structure has become increasingly important. Reports regarding AE-related costs commonly analyze events identified from computerized medical records, payment claims with diagnosis coding, or global trigger tools, while analyses have usually focused on one or a few event categories, such as falls.
Among the methods used to analyze PSE-related costs [
17], linear regression is the conventional method [
18]. However, this method might ignore the interactions between hospital length of stay (LOS) and the occurrence of events that may have a reciprocal causal relationship. Propensity score-matching (PSM) with comparisons has also been advocated [
19], which provides insight into cohort observations under randomization and includes time as an essential factor. Nevertheless, reports applying and comparing PSM with other methods in analyzing PSE-related costs are limited. The purpose of this study was to apply these two methods to analyze the healthcare costs and utilization associated with the occurrence of inpatient PSEs.
The remainder of the paper is organized as follows. The next section outlines the methodology of this study. We subsequently present and discuss our findings and provide concluding remarks.
2. Materials and Methods
2.1. Study Design and Setting
This retrospective study was conducted at National Taiwan University Hospital to analyze the cost and healthcare utilization associated with PSEs between 1 January 2010, and 31 December 2015. The Institutional Research Ethics Committee of the National Taiwan University Hospital approved this study (#201709046RINA) and waived the need for informed consent from the patients.
The hospital is a medical center, with about 2300 beds and approximately 6800 workers, including 460 attending physicians, 650 residents, 2800 nurses, and other workers. The hospital’s Incident Reporting System (IRS) was established in 2002; all of the PSE reports in this study were provided electronically. Staff from the Center for Quality Management of the hospital validated the content, patient harm, and severity of each PSE after receiving the report.
2.2. Participants
Patients admitted to the hospital during the study period were included. Patients with at least one reported PSE were labeled as PSE cases; the remaining hospitalizations without any reported PSE were classified as controls. This labeling was used to retrieve relevant clinical, administrative and claims data upon request from the Integrated Medical Database (NTUH-iMD) of the Department of Research of the National Taiwan University Hospital, who provided de-identified data for research purposes. Patients were excluded if they had at least one of the following conditions: (1) admitted for end-of-life management; (2) admission not paid by the National Health Insurance program.
A
patient safety event (PSE) was defined as an unexpected or unintended event, which could have led to or did lead to harm to the involved patient. These events consisted of the following types: an
adverse event (AE) was defined as an injury caused during the health care process rather than by the underlying disease or condition of the patient; a
no-harm event was defined as an event which resulted in no harm to the patient, or the effect was so minor that the patient could not even feel it; a
near miss event was defined as an event that may have caused an accident, injury or illness, but did not due to an unintentional or timely intervention [
20].
2.3. Variables
To analyze the association of PSEs with costs and healthcare utilization, the investigators acquired the following variable: demographic data, admitting department, category of event, the severity of harm to the patient, claimed medical expenses of the hospitalization, and length of stay.
2.4. Data Sources, Collection, and Measurements
The following data were collected for each PSE from the IRS database: age of the patient, department, and ward for admission, category of events resulting in the PSE, severity of harm to the patient. Furthermore, after the approval by the Research Ethics Committee, the study investigators requested the Integrated Medical Database (NTUH-iMD) of the Department of Research of this hospital to providing de-identified clinical and administrative data. The database provided cases with de-identification data for research purposes. Patient-level costs were obtained from the claims for each admission for the reimbursement by the National Health Insurance of the Taiwan, which provided universal healthcare coverage of the patients and set the prices for hospital care. Costs were converted into US dollars based on the exchange rate toward the end of 2015. The inflation rate in Taiwan has been maintained at a steady low level, about 1.09% annually; therefore, the authors decided that the inflation rate would not be included in the adjustment. The primary outcome measures were total financial costs based on the claims and length of stay.
2.5. Bias and Study Size
Since the under-reporting of PSEs remains a commonly understood issue across the institution, the study might encounter selection bias to assigning patients with PSE to the control group because of the non-reporting of events. This situation could also result in information bias, in that the control group could also suffer from unreported PSEs. These types of bias could result in underestimation of the impact of costs and LOS caused by the PSEs.
As the investigators were not able to have a reference cost estimate from a pilot analysis in the same hospital or a similar study from another institution, the study size was not estimated, and the investigators decided that all of the reported PSE in the IRS would be analyzed after proper inclusion and exclusion processes were applied.
2.6. Statistical Analysis
We first performed a descriptive analysis of the included reported PSEs regarding their category, severity, and time of occurrence concerning the admission date to show the general background scenario. Categorical data are expressed as number (%). Continuous variables were expressed as mean ± SD and 95% confidence interval (95% CI). We then performed inference analysis to compare the direct costs and LOS between the cases and controls according to the model used. We then selected the most common categories of PSE, together contributing to at least 50% of the PSEs. In this study, we used two models to estimate differences in the direct costs and LOS between the cases and controls. In both models, we only selected the cases with a reported PSE within the first 14 days of hospitalization as the cases and then stratified these cases into 14 groups based on the day when the PSE was reported. The control cases were those who did not have any reported PSE and were not discharged until at least the stratified day. For example, for the reported PSEs that occurred on the third day after admission, we only included controls with an LOS of at least three days.
In the first model, we performed linear regression analysis to estimate direct costs and LOS, including age, gender, diagnosis, Charlson comorbidity index [
21,
22], and the main department where the care was provided. In the second model, the propensity score was the conditional probability for having a PSE as a binary dependent variable. Age, gender, main diagnosis, Charlson comorbidity index, admission department and calendar year of admission were added into a non-parsimonious multivariable logistic regression model to predict the effect of PSEs. The propensity score for each individual was the predicted probability derived from the logistic equation. We used PSM to compare the costs and LOS between the groups with and without PSEs. The estimated propensity scores from the datasets were then combined using the mean of the individual estimates in each dataset, according to Rubin’s rule [
23]. One-to-one matching by propensity score, without replacement, was performed using the nearest neighbor method within a caliber of 0.1. We performed matching separately within each event category. PSM was performed using SAS software (version 9.4, SAS Institute Inc., SAS Campus Drive, Cary, North Carolina, USA). The matched cases were then identified, and comparisons were made between cases and controls.
Continuous variables were expressed as mean ± SD. Statistical analyses were performed using SAS software (SAS Institute Inc., USA). All statistical comparisons were two-tailed. A p-value < 0.05 was considered to indicate statistical significance.
4. Discussion
Our results showed that of the hospitalizations with and without PSEs during the first 14 days of admission, the extents of differences in costs and LOS varied across the three PSE categories. Differences in costs and LOS for fall events were not significant by either of the two comparison models; however, the differences were more consistently significant for tube and line events and were less consistently significant for medication events by the two estimation models. A feature of this paper is the novel approach with the reference day-stratified grouping of cases against controls based on the occurrence or absence of patient safety events. We believe that this methodology has not been reported in the literature. We also employed two analysis methods, including linear regression and propensity score-matched comparison, and the results were similar between the two methods. The comparison between these two methods based on reference day stratification approach has not been reported before, either.
The comparisons of costs and LOS in this study were based on the presence or absence of reported PSEs through the institutional IRS; thus, in addition to AEs, the analyses also included events without associated patient harm and those reported for identified undesired processes, materials, equipment and facilities related to patient care. Therefore, this study provides a broader approach to the understanding of undesired contexts of care concerning the burden of healthcare systems. A study showed that AE-related additional financial costs were highest for surgery-related events, with an additional 10.9 hospital days and additional costs of USD 57,727, while an event occurring at medical wards led to an additional 9.8 hospital days and an additional cost of USD 38,656 [
24]. Another study reported that AEs resulted in an average increase of 2.2 hospital days and an increased cost of USD 3244, with 31.6% of the events being preventable and contributing to 4.6 of hospital days and USD 5857 of related costs [
25]. Our study, despite a broader approach, including PSEs without patient harm, still showed significant financial and utilization impacts, except for the category of fall events. This finding suggests that PSEs are at least a marker of increased cost and utilization of healthcare for hospitalizations.
Our claims-data-based approach to estimate costs might have ignored other hidden costs such as administration costs in terms of reporting, data collection, meeting and discussion, improvement activities, and maintenance of the facilities and materials related to health care. Although the cost difference was not significant for fall events, the aggregated cost may still have had a large impact on healthcare. However, standardized methods to assess the burden related to PSEs are currently lacking [
17]. Although the data sources studies accessed might include claim database [
26], full cost-benefit/utility evaluations are rarely completed as they are resource-intensive and often require unavailable data [
27]. We used two models to compare differences in costs and LOS, including multivariable linear regression and PSM. This study employed the reference days as the basis of grouping for comparison, which might exert control of these two variables (features) by assuming the PSE cases and control cases had similar costs until the reference day because of the same in-hospital stay. As mentioned in the Introduction, we were worried about the soundness of linear regression that this method might ignore the interactions between hospital length of stay (LOS) and the occurrence of events that may have a reciprocal causal relationship. Therefore, we proposed a reference-day-based PSM comparison to understand the robustness of the model. The variables included in these models were limited; therefore, we were not able to exclude interference by other factors such as the severity of illness and interventions, which may vary widely, despite a large number of control cases in this study. As can be seen in
Table 4,
Table 5 and
Table 6, while the cost differences were not necessarily estimated higher by either LOS model or PSM model, the LOS differences appeared to be estimated to be higher by the PSM model, counting those reference days with statistical significance in the differences. The literature lacked a previous report of the comparison between these two models in similar studies regarding the impact of PSE or AE on costs and LOS; therefore, further studies are required to elucidate whether these two models do show discrepancy in results. Additional experiments have to be carried out to validate the feasibility and robustness of the proposed models.
Given the higher financial burden and longer LOS, our study may not imply causality between the occurrence of PSEs and higher costs. Specific categories of PSE, such as tube and line and medication, would not occur if the patients do not receive the specific materials for the intervention. The complexity of a clinical condition might also be associated with the materials used, and intervention received, contributing to a higher cost and carrying a higher risk of PSE. Therefore, the cost and LOS may have the same risk factors as the cases with PSEs. Moreover, as most of the patients with PSEs did not suffer from any harm, a longer LOS is less likely to be caused by the PSE, but may be related to the disease and the care provided. Another possible explanation is that the occurrence of a PSE, even if it did not result in patient harm, may be linked to an undesired quality relating to the care process and undesired context. This situation may also have affected the effectiveness and efficiency of care, as reflected by increased costs and prolonged LOS. Further studies to comprehensively investigate the financial impact associated with these events are needed.
Limitations
There are several limitations to this study. First, reporting PSE was voluntary; therefore, underreporting of PSEs is highly probable, as has been reported previously [
28]. However, we do not know whether this led to an over- or underestimation of the associated costs and LOS. Supplementary methods, such as trigger tools and audits, may be useful [
29]. Second, not all of the PSEs may have been related to patient harm, such as equipment defects and undesired working conditions, and the actual process of care may not have been affected. Third, estimations were not adjusted for disease severity; therefore, we did not know whether the patients who experienced PSEs were more ill than the control group patients and whether that also contributed to higher medical expenses related to the disease and routine care rather than to the events themselves. Fourth, we did not use purchasing power parity for converting the costs to US dollars, which might be more appropriate than exchange rates. Fifth, we chose not to compare costs and LOS between cases with patient harm and those without PSEs, mainly because the number of cases with patient harm was too small, with less than one-third of cases involving harm and less than 5% moderate or more severe harm. The number of cases would have been even smaller if they were separated by the day of occurrence, and this may have reduced the statistical power of the analyses. Larger scale analysis is warranted. Based on these limitations, we hypothesize that the negative impact of financial costs on healthcare systems may be much higher than estimated in this study. We believe that the occurrence of PSEs during admission is a marker of high cost for care services. Last, our study chose the reference-day-stratified comparison of groups of patients without any PSE before the reference day. Therefore, competing risk analyses were not performed in this study, and we do not know if patients with repeated PSEs of the same or different event category had a similar pattern of cost and LOS differences. Our findings, therefore, require further validation.