Customize your JAMA Network experience by selecting one or more topics from the list below.
Amarasingham R, Plantinga L, Diener-West M, Gaskin DJ, Powe NR. Clinical Information Technologies and Inpatient Outcomes: A Multiple Hospital Study. Arch Intern Med. 2009;169(2):108–114. doi:10.1001/archinternmed.2008.520
Copyright 2009 American Medical Association. All Rights Reserved. Applicable FARS/DFARS Restrictions Apply to Government Use.2009
Despite speculation that clinical information technologies will improve clinical and financial outcomes, few studies have examined this relationship in a large number of hospitals.
We conducted a cross-sectional study of urban hospitals in Texas using the Clinical Information Technology Assessment Tool, which measures a hospital's level of automation based on physician interactions with the information system. After adjustment for potential confounders, we examined whether greater automation of hospital information was associated with reduced rates of inpatient mortality, complications, costs, and length of stay for 167 233 patients older than 50 years admitted to responding hospitals between December 1, 2005, and May 30, 2006.
We received a sufficient number of responses from 41 of 72 hospitals (58%). For all medical conditions studied, a 10-point increase in the automation of notes and records was associated with a 15% decrease in the adjusted odds of fatal hospitalizations (0.85; 95% confidence interval, 0.74-0.97). Higher scores in order entry were associated with 9% and 55% decreases in the adjusted odds of death for myocardial infarction and coronary artery bypass graft procedures, respectively. For all causes of hospitalization, higher scores in decision support were associated with a 16% decrease in the adjusted odds of complications (0.84; 95% confidence interval, 0.79-0.90). Higher scores on test results, order entry, and decision support were associated with lower costs for all hospital admissions (−$110, −$132, and −$538, respectively; P < .05).
Hospitals with automated notes and records, order entry, and clinical decision support had fewer complications, lower mortality rates, and lower costs.
In recent years, American health care has been criticized as fragmented, expensive, unsafe, and unfair.1 Clinical or “health” information technologies, such as electronic medical records, computerized provider order entry systems, and clinical decision support systems, have emerged as one antidote, promising reductions in waste, gains in communication, improvements in quality, and new accountabilities through automated performance measurement. Benefits have emerged.2,3 However, studies examining the impact of these technologies are not easily generalized; most studies are limited to single-site evaluations, often academic hospitals that have developed their systems internally and incrementally, sometimes for decades.3 In contrast, most US hospitals must consider purchasing commercially developed information systems with a broad range of electronic capabilities. Few studies have been performed across multiple hospitals to forecast the effect of clinical information technologies in these settings.
The clinical information system of a hospital can be divided into 4 principal subdomains: notes and records, test results, order entry, and decision support. Information in each of these areas would ordinarily be managed through paper-based systems; to the degree that a hospital is “paperless,” these functions are automated. We previously developed a physician-based assessment tool that quantifies the degree to which a hospital has effectively computerized these 4 subdomains. The instrument has demonstrated reliability and validity.4 In this study, we examined the association between a hospital's automation and inpatient mortality, complications, costs, and length of stay (LOS) among patients with 4 medical conditions in a diverse group of Texas hospitals. It has been argued that automating the various domains of a hospital's clinical information system will improve communication among health care providers, quicken responses to abnormal diagnostic test results, enhance clinical decision making, and improve adherence to guidelines. The 4 medical conditions included in this study—myocardial infarction, congestive heart failure, coronary artery bypass grafting, and pneumonia—are common and are thought to be sensitive to clinical guidelines. We hypothesized that greater levels of automation would be associated with better outcomes.
We conducted a cross-sectional study of urban hospitals in Texas. We sampled from 72 general acute-care hospitals located within 10 geographically dispersed metropolitan statistical areas in Texas: Abilene, Austin, Dallas, El Paso, Houston, Laredo, Lubbock, McAllen, San Angelo, and San Antonio. We selected Texas as the site for study because it contains a large and diverse patient population and a wide range of hospitals for which specific clinical outcomes could be obtained. We excluded pediatric, specialty, or long-term care hospitals; hospitals that were in the process of closing or merging with another facility; and hospitals for which we could not obtain discharge data for the targeted diagnosis related groups. The Johns Hopkins University School of Medicine Institutional Review Board approved the research protocol.
We measured automation in the hospitals using the Clinical Information Technology Assessment Tool (CITAT), which is a questionnaire-based tool administered to physicians who provide inpatient care. The CITAT was produced in 8 steps according to rigorous methods of instrument development. This instrument has been tested and validated in 4 US hospitals and demonstrates discriminant validity, convergent validity, reliability, and precision.4,5 The CITAT assesses a system's automation, or the degree to which clinical information processes in the hospital are computerized. Automation is divided into the 4 subdomains: test results, notes and records, order entry, and decision support.4The instrument contains several items for each information technology subdomain; 3 factors are required to achieve a high score on any individual item: the information process must be available as a fully computerized process; the physician must know how to activate the computerized process; and he or she must choose the computerized process over other alternatives, such as using paper-based documentation, making a telephone call, or referring to a clinical textbook. These subdomain scores, scaled from 0 (lowest possible score) to 100 (maximum possible score), served as independent variables in this study.
Using the American Medical Association Physician Masterfile, we selected a 50% random sample of physicians from those who had practice locations in the designated metropolitan statistical areas and who practiced internal medicine (including 9 subspecialties), general surgery (including 10 subspecialties), or family practice (n = 7432). We mailed surveys to each of the selected physicians between December 1, 2005, and May 30, 2006. Physicians indicated whether they practiced inpatient medicine and, if so, selected the hospital in which they provided most of their inpatient care. To be eligible, physicians had to actively practice in 1 of 72 hospitals selected for this study. As guided by prior work,5 hospitals for which we did not receive 5 randomly sampled physician responses were eliminated from further analysis owing to the possibility of unstable estimates. For each respondent, 4 separate subdomain scores were calculated using previously described methods.4,5 Each hospital was then assigned the median value of the scores derived from respondents affiliated with that hospital.
We examined inpatient mortality, complications, costs, and LOS among patients older than 50 years who were admitted between December 1, 2005, and May 30, 2006, at any of the 72 study hospitals. Discharge-level data for 167 233 patients, including information on all 4 outcomes, were obtained from a hospital claims data file (provided by the Texas Hospital Association and compiled by Solucient, Evanston, Illinois) and merged with the CITAT data. Total costs were derived from charges using hospital-specific cost to charge ratios for each discharge.
Hospital characteristics were obtained from the 2005 survey of the Texas Hospital Association and the American Hospital Association annual survey of Texas hospitals. For each hospital in our sample, we obtained the ownership status (public, private/nonprofit, and private/for-profit), bed size (number of beds), and total margin. Estimates of the risks of complication and mortality for each hospitalization were obtained from the risk-adjusted complication index and risk-adjusted mortality index variables provided in the hospital claims data file. These previously validated indices adjust for severity and case-mix differences based on age, sex, principal diagnoses, and procedures performed and are stratified by similar hospitals (ie, number of beds, urban vs rural, teaching status, and census tract).6,7 The hospital-level data (CITAT scores and hospital characteristics) were linked by hospital to the discharge data.
We compared the characteristics of hospitals with 5 or more physician responses with excluded hospitals using Pearson χ2 and t tests for categorical and continuous variables, respectively. We examined the association between each hospital's automation subdomain scores with each of the 4 dependent variables (mortality, complications, costs, and LOS). Independent variables were interpreted using 10-point increments. For each of the dependent variables, analyses were performed for all discharges and for those discharges with the following principal International Classification of Diseases, Ninth Revision, codes: myocardial infarction (410.xx), heart failure (428.xx, 398.91, 402.01, 402.11, 402.91, 404.00, 404.01, 404.03, 404.10, 404.11, 404.13, 404.90, 404.91, and 404.93), coronary artery bypass graft (36.10-36.17 and 36.19), and pneumonia (480.0-483.8, 485-486, and 487.0).8
Logistic regression was used to estimate odds ratios (ORs) for death or complication for each 10-point increase in the automation and subdomain score. Costs and LOS were examined with linear regression after log transformation and retransformed by the Duan smearing method for presentation.9 Multivariable adjustment included risk (risk-adjusted complication and mortality indices) and hospital characteristics associated with the independent variables. We accounted for possible within-hospital clustering of patient outcomes in all discharge-level analyses by obtaining robust variance-covariance matrix estimates for all models.10,11 When there was a statistically significant relationship between the dependent and independent variables, we performed a sensitivity analysis to test the linearity assumptions; subdomain scores were divided into tertiles. Statistical significance was set at P ≤ .05. Because of the possibility of false-positive relationships owing to multiple hypothesis testing, a Bonferroni correction was performed as an additional sensitivity analysis to adjust for the investigation of 4 automation subdomains for each disease condition and outcome. STATA statistical software, version 9.2, (STATA Corp, College Station, Texas) was used for all analyses.
We received 5 or more physician responses for 41 of 72 targeted hospitals (58% response rate; mean number of responses, 9) (Table 1). There were no statistically significant differences in hospital ownership, operating margin, total margin, safety net status, or information technology operating expenses between responding and nonresponding hospitals. However, responding hospitals tended to be larger (mean number of beds, 402 vs 216 for nonresponders; P = .001) and more academic (6 teaching hospitals [15%] vs 0 teaching hospitals, respectively; P = .03). There were no statistically significant differences between responders and nonresponders with respect to their hospital mortality rate (P = .41), complication rate (P = .86), or median costs (P = .14). Hospitals for which we did not receive a response had a longer median LOS (3 days longer; P = .04).
Overall, hospitals scored low on most CITAT subdomains; higher scores were observed for the test results and notes and records subdomains (mean scores, 50.2 and 28.5, respectively) compared with order entry and decision support (3.7 and 2.6, respectively).
Across a variety of clinical conditions, higher CITAT scores were associated with decreased adjusted ORs for fatal hospitalizations (Table 2). Higher notes and records scores were associated with a statistically significant decrease in the adjusted odds of inpatient mortality in all-cause hospitalizations (OR, 0.85; 95% confidence interval [CI], 0.74-0.97). Hospitals with higher order entry scores were associated with decreased adjusted odds for fatal hospitalizations for patients admitted with myocardial infarction (0.91; 0.83-0.99) and coronary artery bypass graft procedures (0.45; 0.29-0.68). Higher decision support scores were associated with decreased adjusted odds for mortality owing to pneumonia (0.79; 0.63-1.00). Generally, patterns were internally consistent across disease conditions and automation subdomain regardless of statistical significance.
Of the automation subdomains, a higher decision support score was consistently associated with decreased adjusted odds for complications (Table 3). These results were statistically significant for all causes (OR, 0.84; 95% CI, 0.79-0.90) and myocardial infarction (0.63; 0.45-0.87). Contrary to this trend, we observed that a higher notes and records score was associated with increased adjusted odds for complications associated with heart failure (1.35; 1.16-1.57).
For nearly all clinical conditions, higher scores on decision support, order entry, and test results were associated with lower mean hospital costs (Table 4). Higher test results scores were statistically significantly associated with lower adjusted costs for all hospital admissions (OR, −$110; 95% CI, −$181 to −$20) and for heart failure (−$207; −$272 to −$128). A higher order entry score was associated with statistically significantly lower adjusted costs for all conditions (−$132; −$232 to −$13). As with test results and order entry, a higher decision support score was also associated with lower adjusted costs for all conditions (−$538; −$704 to −$333) and for coronary artery bypass grafting (−$1043; −$1729 to −$55).
No clear pattern emerged in the relationship between CITAT scores and hospital LOS by clinical condition (Table 5). A 10-point increase in score for order entry and decision support was associated with decreased but not clinically meaningful LOS for heart failure (−0.09 and −0.22 days, respectively). In all cases, differences in LOS were modest in either direction.
All adjusted and statistically significant associations in the primary analysis retained the direction of their association when the independent variable was categorized into tertiles, but fewer were statistically significant (P < .05 for 7 of 22 associations). As in the primary analysis, higher notes and records scores were associated with lower adjusted rates of mortality (1.9%, 1.6%, and 1.4% in the low, intermediate, and high tertiles, respectively; P = .02 for the trend); higher scores for decision support were associated with lower rates of complications (4.1%, 3.8%, and 3.5%, respectively; P = .045 for the trend); and higher scores for decision support were associated with lower median costs ($5067, $4966, and $4498, respectively; P = .008 for the trend). Similar trends were observed for the relationships between tests results and costs, order entry and costs, and order entry and LOS, which mirrored the primary analysis.
As a separate sensitivity analysis, we performed a Bonferroni correction to account for the investigation of the 4 automation subdomains. Most associations remained statistically significant at the adjusted significance level (P < .0125, the overall significance level of .05 divided by 4).
This study provides empirical evidence that greater automation of a hospital's information system may be associated with reductions in mortality, complications, and costs. Higher decision support scores were associated with statistically significant reductions in the odds of complications among all causes and for myocardial infarction specifically, and with reductions in the odds of death for pneumonia. Among the remaining associations that were not statistically significant, all showed trends toward reductions in mortality and complications. Prior reports have suggested that decision support helps health care providers manage large amounts of incoming data, provides context for decision making in light of guidelines, and may help physicians avoid “sins of omission,” reputed by some authors to be the largest source of medical errors.12-16 Knowledge aids provided in this fashion could reduce the risk of complications and possibly death, providing a theoretical basis for the association we observed.
Higher order entry scores were associated with reductions in the odds of death for myocardial infarction and coronary artery bypass graft surgery. Prior studies of order entry at single institutions have shown mixed results with respect to mortality17,18; nevertheless, salutary effects have been observed for other end points, including reductions in the number of adverse drug events, improved legibility of orders, and fewer callbacks to ordering physicians.19 These factors may mediate the reductions in mortality odds we observe with order entry. Some have suggested that the use of information technologies in the clinical environment poses certain risks, noted as “e-iatrogenesis.”20 For example, previous studies have raised concerns that problems in information representation within computerized order entry systems could facilitate errors.21,22 We found no relationship between the degree to which physicians' orders were computerized and the rate of complications.
Greater automation of test results was not associated with a lower risk of hospital death or complications. With respect to test results, it may be that a minimal level of automation is sufficient to protect patients. Given that a high proportion of hospitals have already automated test results, this minimal level is likely in place for most institutions. Advanced automated test result features, such as user customization or better search and retrieval capabilities, could increase physician satisfaction but may have little effect on patient outcomes.
Our results suggest that the relationship between outcomes and automated notes and records is nuanced but internally consistent. Each 10-point increase in the notes and records score was associated with a 15% reduction in odds of inpatient death for all causes. Hospitals in the highest tertile of the notes and records subdomain score had a 1.4% adjusted rate of mortality, compared with a 1.9% adjusted rate among hospitals in the lowest tertile. This would suggest that for every 1000 patients, 5 fewer patients die at hospitals with the highest notes and records scores. Higher scores on the notes and record subdomain were also associated with reduced odds of death for each of the 4 individual conditions examined in the mortality analysis, although none were statistically significant. Smaller studies suggest that electronic documentation allows clinicians faster and more complete access to the patient record, improves communication among health care providers, and enhances the contributions of supervising physicians.23 In contrast, increases in the notes and records score were associated with statistically significant increases in the odds of complications for heart failure. Although this may be of concern, another explanation might be that a higher complication rate simply reflects an improved capability to identify adverse events through electronic documentation. A similar mechanism may underlie the relationship between notes and records and costs. The remaining associations tested in the complication analyses retained the direction noted with heart failure, although no others were statistically significant.
Higher scores on test results, order entry, and decision support were overwhelmingly associated with lower hospital costs. Of the 15 associations tested in these categories, 14 demonstrated an inverse relationship between the information technology score and total costs and 10 of these were statistically significant. The associated reduction in costs for some conditions was substantial. Relationships are less clear with respect to LOS, and any effect is modest. Because LOS has decreased substantially during the last several decades, in part because of increased scrutiny by payers, this measure may already be so low as to be resistant to the efficiencies introduced by information technology.24 However, hospitals for which we had responses had shorter LOS, on average, than did nonresponders. It is possible that for hospitals with longer LOS at baseline, the effect of clinical information technologies might be more profound.
This study has a number of strengths. Our approach incorporates what some have described as the sociotechnical environment of the clinical workplace. This view holds that a successful information technology implementation jointly optimizes the technology and the social aspects of an organization (eg, its policies, values, norms, and culture).25 To properly account for the impact of clinical information technologies in a way that can be replicated, the information technology variable must be measured in the context of the sociotechnical environment in which it is implemented. This study measures a hospital's level of automation based on physicians' daily interaction with the information system, avoiding simple terminological definitions that may not account for usage, maturation, and capabilities of the information system. If there is insufficient user training, if the technology itself is unfriendly, or if the physician and organizational routines are not aligned with the technology, the information technology score for that hospital will be low, regardless of the cost or scope of the technologic acquisition. This study also includes more hospitals and hospitals of greater organizational variety than prior studies, overcoming criticisms that a small number of specific academic hospitals are overrepresented in examinations of clinical information technology.3 Finally, our results are congruent with recent studies suggesting that the adoption of clinical information technologies remains low but follows certain patterns.26,27 Our findings are consistent with these patterns, lending our methods, and measurement tool, an independent measure of validity. For example, the computerized display of laboratory results has been among the first aspects to be automated.27 In the last decade, digitization of radiological images has also increased.27 Both of these components fall under the test results subdomain, which in our study showed the greatest degree of adoption. Electronic decision support is perhaps the most challenging component to implement because it requires all other components first. Our results, in which scores for notes and records are higher than order entry and decision support, are consistent with this pattern of adoption.
Despite these strengths, this study has important limitations. First, the design does not consider a number of organizational confounders that could explain superior clinical outcomes, most notably a hospital's emphasis on safety and quality. We adjusted for structural variables that could be related to quality, including number of beds, total margin, and hospital ownership (in our sample, teaching status and number of beds were not associated with the independent variables). Hospitals that do not emphasize a culture of safety or encourage continuous improvement are probably incapable of replicating the type of sociotechnical environment that would produce a high score in this study. In this sense, the distinction between whether a hospital's superior outcomes are because of its emphasis on quality or its investment in clinical information technologies becomes less meaningful because both are likely required to produce a high-functioning sociotechnical environment. Future studies using mixed or qualitative methods at several representative sites may help clarify these relationships.
Second, our analysis explored a number of information technology functions, raising issues of multiple hypothesis testing and the possibility of some false-positive relationships. As with all cross-sectional studies, positive associations will need to be confirmed in repeated studies. Nevertheless, the consistency of our findings suggests that the observed patterns are real. These results may also indicate that, because disease conditions are unique, the factors resulting in better outcomes for these conditions are themselves unique. Indeed, the lack of statistical significance among certain associations may simply indicate that clinical information technology is not a panacea for all disease conditions. Third, it is important to recognize that the observed associations between improved outcomes and information technology subdomains can only be extrapolated for the observed range of CITAT scores in our sample. As we observed with test results, it is possible that at higher ranges of order entry and decision support, differences among hospital outcomes will diminish. Fourth, although we obtained a satisfactory response rate (58%), we experienced higher response rates among larger hospitals, academic hospitals, and hospitals with shorter LOS. This may limit the application of our findings.
Clinical information technologies hold great promise as a tool to improve hospital medicine. We found that, for certain conditions, greater automation of a hospital's information system may be associated with reductions in mortality, complications, and costs, suggesting that information technologies that are properly designed and executed around clinical workflows could meet that promise.
Correspondence: Ruben Amarasingham, MD, MBA, Center for Knowledge Translation and Clinical Innovation, Parkland Health & Hospital System, 5201 Harry Hines Blvd, Dallas, TX 75235 (email@example.com).
Accepted for Publication: June 27, 2008.
Author Contributions: Drs Amarasingham and Powe had full access to all of the data in the study and take responsibility for the integrity of the data and the accuracy of the data analysis. Study concept and design: Amarasingham and Powe. Acquisition of data: Amarasingham, Plantinga, and Powe. Analysis and interpretation of data: Amarasingham, Plantinga, Diener-West, Gaskin, and Powe. Drafting of the manuscript: Amarasingham and Powe. Critical revisions of the manuscript: Amarasingham, Plantinga, Diener-West, Gaskin, and Powe. Statistical analysis: Amarasingham, Plantinga, Diener-West, Gaskin, and Powe. Obtained funding: Amarasingham and Powe. Administrative, technical, or material support: Amarasingham and Powe. Study supervision: Amarasingham and Powe.
Financial Disclosure: None reported.
Funding/Support: This study was supported by grant 20050277 from the Commonwealth Fund, New York, New York.
Disclaimer: The Commonwealth Fund was not involved in the design and conduct of the study; the collection, management, analysis, or interpretation of data; or the preparation, review, or approval of the manuscript.
Additional Contributions: Aaron Cunningham, BA, served as a research assistant for this project and was paid for his assistance with survey preparation, data collection, and data entry. We thank the Commonwealth Fund for its generous funding support. We also thank the Texas Health Institute, the Texas Medical Association, the Texas Hospital Association, the TMF Health Quality Institute, and the Texas Department of State Health Services for their help facilitating the study in Texas.
Create a personal account or sign in to: