Modifying ICD-9-CM coding of secondary diagnoses to improve risk-adjustment of inpatient mortality rates

Use of Deep Learning to Identify Peripheral Arterial Disease Cases From Narrative Clinical Notes

INTRODUCTION

Peripheral arterial disease (PAD) is the leading cause of amputation in the United States. Despite affecting 8.5 million Americans and more than 200 million people globally, there are significant gaps in awareness by both patients and providers. Ongoing efforts to raise PAD awareness among both the public and health-care professionals have not met widespread success. Thus, there is a need for alternative methods for identifying PAD patients. One potentially promising strategy leverages natural language processing (NLP) to digitally screen patients for PAD. Prior approaches have applied keyword search (KWS) to billing codes or unstructured clinical narratives to identify patients with PAD. However, KWS is limited by its lack of flexibility, the need for manual algorithm development, inconsistent validation, and an inherent failure to capture patients with undiagnosed PAD. Recent advances in deep learning (DL) allow modern NLP models to learn a conceptual representation of the verbiage associated with PAD. This capability may overcome the characteristic constraints of applying strict rule-based algorithms (i.e., searching for a disease-defining set of keywords or billing codes) to real-world clinical data. Herein, we investigate the use of DL to identify patients with PAD from unstructured notes in the electronic health record (EHR).

METHODS

Using EHR data from a statewide health information exchange, we first created a dataset of all patients with diagnostic or procedural codes (International Classification of Diseases version 9 or 10 or Current Procedural Terminology) for PAD. This study population was then subdivided into training (70%) and testing (30%) cohorts. We based ground truth labels (PAD versus no PAD) on the presence of a primary diagnostic or procedural billing code for PAD at the encounter level. We implemented our KWS-based identification strategy using the currently published state-of-the-art algorithm for identifying PAD cases from unstructured EHR data. We developed a DL model using a BioMed-RoBERTa base that was fine-tuned on the training cohort. We compared the performance of the KWS algorithm to our DL model on a binary classification task (PAD versus no PAD).

RESULTS

Our study included 484,363 encounters across 71,355 patients represented in 2,268,062 notes. For the task of correctly identifying PAD related notes in our testing set, the DL outperformed KWS on all model performance measures (Sens 0.70 versus 0.62; Spec 0.99 versus 0.94; PPV 0.82 versus 0.69; NPV 0.97 versus 0.96; Accuracy 0.96 versus 0.91; P value for all comparisons <0.001).

CONCLUSIONS

Our findings suggest that DL outperforms KWS for identifying PAD cases from clinical narratives. Future planned work derived from this project will develop models to stage patients based on clinical scoring systems.

Convergent Validity of 2 Widely Used Methodologies for Calculating the Hospital Standardized Mortality Ratio in Flanders, Belgium

OBJECTIVES

To assess their construct validity, we compared results from 2 models used for estimating hospital standardized mortality ratios (HSMRs) in Belgium. The method of the Flemish Hospital Network (FHN) is based on a logistic regression for each of the 64 All Patient Refined Diagnosis-Related Groups that explain 80% of mortality and uses the Elixhauser score to correct for comorbidities. (H)SMRs published on the 3M-Benchmark-Portal are calculated by a simpler indirect standardization for All Patient Refined Diagnosis-Related Groups and risk of mortality (ROM) at discharge.

METHODS

We used administrative data from all eligible hospital admissions in 22 Flemish hospitals between 2016 and 2019 (FHN, n = 682,935; 3M, n = 2,122,305). We evaluated model discrimination and accuracy and assessed agreement in estimated HSMRs between methods.

RESULTS

The Spearman correlation between HSMRs generated by the FHN model and the standard 3M model was 0.79. Although 2 of 22 hospitals showed opposite classification results, that is, an HSMR significantly <1 according to the FHN method but significantly >1 according to the 3M model, classification agreement between methods was significant (agreement for 59.1% of hospitals, κ = 0.45). The 3M model ( c statistic = 0.96, adjusted Brier score = 26%) outperformed the FHN model (0.87, 17%). However, using ROM at admission instead of at discharge in the 3M model significantly reduced model performance ( c statistic = 0.94, adjusted Brier score = 21%), but yielded similar HSMR estimates and eliminated part of the discrepancy with FHN results.

CONCLUSIONS

Results of both models agreed relatively well, supporting convergent validity. Whereas the FHN method only adjusts for disease severity at admission, the ROM indicator of the 3M model includes diagnoses not present on admission. Although diagnosis codes generated by complications during hospitalization have the tendency to increase the predictive performance of a model, these should not be included in risk adjustment procedures.

Using structured pathology data to predict hospital-wide mortality at admission

Early prediction of in-hospital mortality can improve patient outcome. Current prediction models for in-hospital mortality focus mainly on specific pathologies. Structured pathology data is hospital-wide readily available and is primarily used for e.g. financing purposes. We aim to build a predictive model at admission using the International Classification of Diseases (ICD) codes as predictors and investigate the effect of the self-evident DNR (“Do Not Resuscitate”) diagnosis codes and palliative care codes. We compare the models using ICD-10-CM codes with Risk of Mortality (RoM) and Charlson Comorbidity Index (CCI) as predictors using the Random Forests modeling approach. We use the Present on Admission flag to distinguish which diagnoses are present on admission. The study is performed in a single center (Ghent University Hospital) with the inclusion of 36 368 patients, all discharged in 2017. Our model at admission using ICD-10-CM codes (AUCROC = 0.9477) outperforms the model using RoM (AUCROC = 0.8797 and CCI (AUCROC = 0.7435). We confirmed that DNR and palliative care codes have a strong impact on the model resulting in a decrease of 7% for the ICD model (AUCROC = 0.8791) at admission. We therefore conclude that a model with a sufficient predictive performance can be derived from structured pathology data, and if real-time available, can serve as a prerequisite to develop a practical clinical decision support system for physicians.

Comparing Standard Performance and Outcome Measures in Hospitalized Pituitary Tumor Patients with Secretory versus Nonsecretory Tumors

BACKGROUND

Patient safety indicators (PSIs) and hospital-acquired conditions (HACs) are reported quality measures. We compared their prevalence in patients with secretory and nonsecretory pituitary adenoma using the National (Nationwide) Inpatient Sample (NIS), Healthcare Cost and Utilization Project, Agency for Healthcare Research and Quality.

METHODS

The NIS was queried for hospitalizations 2002-2014 involving pituitary adenomas. Prevalence of PSI, HAC, and 9 pituitary-related complications was determined using International Classification of Diseases, Ninth Revision codes. Patient risk factors were evaluated through multivariate analysis.

RESULTS

A total of 20,743 patients with nonsecretory tumor and 3385 patients with secretory tumor were identified. Among patients with nonsecretory tumor, 3.79% experienced any PSI or HAC. Of patients with secretory tumor, 2.54% had any PSI or HAC. Before adjusting for covariation, secretory patients were less likely to have any PSI or HAC (odds ratio [OR], 0.652; P = 0.0002), experience any pituitary-related complication (OR, 0.804; P < 0.0001), have a poor outcome (hazard ratio [HR], 0.435; P < 0.0001), and die during hospitalization (HR, 0.293; P = 0.0015). Secretory patients had significantly shorter mean hospital length of stay (secretory/nonsecretory percent difference, -11.95%; P < 0.0001). However, inverse propensity score-weighted ORs comparing the groups' outcomes showed that there was no significant difference in the prevalence of any PSIs and HACs (OR, 0.963; P = 0.8570), pituitary-related complications (OR, 0.894; P = 0.1321), poor outcomes (HR, 0.990; P = 0.9287), in-hospital death (HR, 0.663; P = 0.2967), and length of stay (percent difference, -2.31%; P = 0.2967) between groups.

CONCLUSIONS

Lack of significant difference in outcome measures after controlling for covariation is consistent with our finding that patients with nonsecretory tumor have more comorbidities on presentation for treatment. PSIs and HACs have limited ability to measure complications specific to pituitary tumors.

Scalable and accurate deep learning with electronic health records

Predictive modeling with electronic health record (EHR) data is anticipated to drive personalized medicine and improve healthcare quality. Constructing predictive statistical models typically requires extraction of curated predictor variables from normalized EHR data, a labor-intensive process that discards the vast majority of information in each patient's record. We propose a representation of patients' entire raw EHR records based on the Fast Healthcare Interoperability Resources (FHIR) format. We demonstrate that deep learning methods using this representation are capable of accurately predicting multiple medical events from multiple centers without site-specific data harmonization. We validated our approach using de-identified EHR data from two US academic medical centers with 216,221 adult patients hospitalized for at least 24 h. In the sequential format we propose, this volume of EHR data unrolled into a total of 46,864,534,945 data points, including clinical notes. Deep learning models achieved high accuracy for tasks such as predicting: in-hospital mortality (area under the receiver operator curve [AUROC] across sites 0.93-0.94), 30-day unplanned readmission (AUROC 0.75-0.76), prolonged length of stay (AUROC 0.85-0.86), and all of a patient's final discharge diagnoses (frequency-weighted AUROC 0.90). These models outperformed traditional, clinically-used predictive models in all cases. We believe that this approach can be used to create accurate and scalable predictions for a variety of clinical scenarios. In a case study of a particular prediction, we demonstrate that neural networks can be used to identify relevant information from the patient's chart.

Appropriateness of ICD-coded Diagnostic Inpatient Hospital Discharge Data for Medical Practice Assessment

Objectives: We performed a systematic review to investigate the quality of diagnostic hospital discharge data (DHDD) in order to gain insight in the usefulness of these data for medical practice assessment. We investigated the methods used to evaluate data quality, factors that determine data quality and its consequences for medical practice assessment.

Methods: We selected studies in which both completeness (or sensitivity: SENS) and correctness (or positive predictive value: PPV) were measured. We used the random-effects model to calculate mean SENS and PPV and to explore the effect of a number of covariates.

Results: The 101 included studies were very heterogeneous. We distinguished six typical study designs. We found a mean SENS of 0.67 (95%CI: 0.62– 0.73) and PPV of 0.76 (95%CI: 0.73– 0.79); SENS was significantly lower for comorbidity and complication studies than for some single disease studies. PPV was significantly higher for Scandinavian countries than for other countries. Recoding compared to re-abstracting of the medical record as a gold standard gave a significantly lower PPV. Diagnostic data were considered appropriate by the authors of the studies for quality of care purposes when both SENS and PPV were at least 0.85. Only 13% of the studies fulfilled this criterion.

Conclusions: Variability in quality of care between settings can easily be overshadowed by variability in data quality. However, the use of DHDD by physicians to evaluate their own medical practice may be useful. But only if physicians are willing to critically interpret the meaning of the information for their medical practice assessment.

Screening entire healthcare system ECG database: Association of deep terminal negativity of P wave in lead V1 and ECG referral with mortality

BACKGROUND

Each encounter of asymptomatic individuals with the healthcare system presents an opportunity for improvement of cardiovascular disease (CVD) awareness and sudden cardiac death (SCD) risk assessment. ECG sign deep terminal negativity of the P wave in V1 (DTNP_V1) was shown to be associated with an increased risk of SCD in the general population.

OBJECTIVE

To evaluate association of DTNP_V1 with all-cause mortality and newly diagnosed atrial fibrillation (AFib) in the large tertiary healthcare system patient population.

METHODS

Retrospective double cohort study compared two levels of exposure (automatically measured amplitude of P-prime (Pp) in V1): DTNP_V1 (Pp from -100μV to -200μV) and ZeroPpV1 (Pp=0). An entire healthcare system (2010-2014) ECG database was screened. Medical records of children and patients with previously diagnosed AFib/atrial flutter (AFl), implanted pacemaker or cardioverter-defibrillator were excluded. DTNP_V1 (n=3,413) and ZeroPpV1 (n=3,405) cohorts were matched by age and sex. Primary outcome was all-cause mortality. Secondary outcomes were newly diagnosed AFib/AFl. Median follow-up was 2.5 y.

RESULTS

DTNP_V1 was associated with all-cause mortality (HR 1.95(1.64-2.31); P<0.0001) and newly diagnosed AFib (HR 1.29(1.04-1.59); P=0.021) after adjustment for CVD, comorbidities, other ECG parameters, medications, and index ECG referral. Index ECG referral by a cardiologist was independently associated with 34% relative risk reduction of mortality (HR 0.66(0.52-0.84); P=0.001), as compared to ECG referral by a non-cardiologist.

CONCLUSION

DTNP_V1 is independently associated with twice higher risk of all-cause death, as compared to patients without P prime in V1. Life-saving effect of the index ECG referral by a cardiologist requires further study.

Can valid and practical risk-prediction or casemix adjustment models, including adjustment for comorbidity, be generated from English hospital administrative data (Hospital Episode Statistics)? A national observational study

Background

NHS hospitals collect a wealth of administrative data covering accident and emergency (A&E) department attendances, inpatient and day case activity, and outpatient appointments. Such data are increasingly being used to compare units and services, but adjusting for risk is difficult.

Objectives

To derive robust risk-adjustment models for various patient groups, including those admitted for heart failure (HF), acute myocardial infarction, colorectal and orthopaedic surgery, and outcomes adjusting for available patient factors such as comorbidity, using England’s Hospital Episode Statistics (HES) data. To assess if more sophisticated statistical methods based on machine learning such as artificial neural networks (ANNs) outperform traditional logistic regression (LR) for risk prediction. To update and assess for the NHS the Charlson index for comorbidity. To assess the usefulness of outpatient data for these models.

Main outcome measures

Mortality, readmission, return to theatre, outpatient non-attendance. For HF patients we considered various readmission measures such as diagnosis-specific and total within a year.

Methods

We systematically reviewed studies comparing two or more comorbidity indices. Logistic regression, ANNs, support vector machines and random forests were compared for mortality and readmission. Models were assessed using discrimination and calibration statistics. Competing risks proportional hazards regression and various count models were used for future admissions and bed-days.

Results

Our systematic review and empirical analysis suggested that for general purposes comorbidity is currently best described by the set of 30 Elixhauser comorbidities plus dementia. Model discrimination was often high for mortality and poor, or at best moderate, for other outcomes, for examplec = 0.62 for readmission andc = 0.73 for death following stroke. Calibration was often good for procedure groups but poorer for diagnosis groups, with overprediction of low risk a common cause. The machine learning methods we investigated offered little beyond LR for their greater complexity and implementation difficulties. For HF, some patient-level predictors differed by primary diagnosis of readmission but not by length of follow-up. Prior non-attendance at outpatient appointments was a useful, strong predictor of readmission. Hospital-level readmission rates for HF did not correlate with readmission rates for non-HF; hospital performance on national audit process measures largely correlated only with HF readmission rates.

Conclusions

Many practical risk-prediction or casemix adjustment models can be generated from HES data using LR, though an extra step is often required for accurate calibration. Including outpatient data in readmission models is useful. The three machine learning methods we assessed added little with these data. Readmission rates for HF patients should be divided by diagnosis on readmission when used for quality improvement.

Future work

As HES data continue to develop and improve in scope and accuracy, they can be used more, for instance A&E records. The return to theatre metric appears promising and could be extended to other index procedures and specialties. While our data did not warrant the testing of a larger number of machine learning methods, databases augmented with physiological and pathology information, for example, might benefit from methods such as boosted trees. Finally, one could apply the HF readmissions analysis to other chronic conditions.

Funding

The National Institute for Health Research Health Services and Delivery Research programme.

Using highly detailed administrative data to predict pneumonia mortality

Background

Mortality prediction models generally require clinical data or are derived from information coded at discharge, limiting adjustment for presenting severity of illness in observational studies using administrative data.

Objectives

To develop and validate a mortality prediction model using administrative data available in the first 2 hospital days.

Research Design

After dividing the dataset into derivation and validation sets, we created a hierarchical generalized linear mortality model that included patient demographics, comorbidities, medications, therapies, and diagnostic tests administered in the first 2 hospital days. We then applied the model to the validation set.

Subjects

Patients aged ≥18 years admitted with pneumonia between July 2007 and June 2010 to 347 hospitals in Premier, Inc.’s Perspective database.

Measures

In hospital mortality.

Results

The derivation cohort included 200,870 patients and the validation cohort had 50,037. Mortality was 7.2%. In the multivariable model, 3 demographic factors, 25 comorbidities, 41 medications, 7 diagnostic tests, and 9 treatments were associated with mortality. Factors that were most strongly associated with mortality included receipt of vasopressors, non-invasive ventilation, and bicarbonate. The model had a c-statistic of 0.85 in both cohorts. In the validation cohort, deciles of predicted risk ranged from 0.3% to 34.3% with observed risk over the same deciles from 0.1% to 33.7%.

Conclusions

A mortality model based on detailed administrative data available in the first 2 hospital days had good discrimination and calibration. The model compares favorably to clinically based prediction models and may be useful in observational studies when clinical data are not available.

[Standards in Medical Informatics: Fundamentals and Applications]

The use of computers in medical practice has enabled novel forms of communication to be developed in health care. The optimization of communication processes is achieved through the use of standards to harmonize the exchange of information and provide a common language for all those involved. This article describes the concept of a standard applied to medical informatics and its importance in the development of various applications, such as computational representation of medical knowledge, disease classification and coding systems, medical literature searches and integration of biological and clinical sciences.

The impact of the present on admission indicator on the accuracy of administrative data for carotid endarterectomy and stenting

BACKGROUND

Administrative data are often hampered by coding errors, absent data, and the difficulty of distinguishing pre-existing conditions from perioperative complications. We evaluated whether the introduction of the present on admission (POA) indicator improved outcome analysis of carotid endarterectomy (CEA) and carotid angioplasty and stenting (CAS) using administrative data.

METHODS

State inpatient databases from California (2005-2008), New York (2008), and New Jersey (2008) were used to identify patients undergoing CAS and CEA. We first analyzed morbidity data without the POA indicator, using International Classification of Diseases, Ninth Revision complication codes (eg, 997.02, iatrogenic cerebrovascular infarction or hemorrhage, postoperative stroke) and diagnosis codes (eg, 433.11, occlusion and stenosis of the carotid artery with cerebral infarction). Then, we applied the POA indicator to both diagnosis and complication codes and calculated the proportion of events that were labeled POA. Symptom status and perioperative stroke rate were compared using these coding approaches.

RESULTS

We identified 21,639 patients who underwent CEA and 3688 patients who underwent CAS. Without the POA indicator, the complication code for stroke indicated a postoperative stroke rate of 1.4% for CEA and 2.4% for CAS. After applying the POA indicator, 54% (CEA) and 62% (CAS) of these strokes were labeled POA. These POA strokes were either preoperative or intraoperative events. Proportion of symptomatic patients ranged from 7% to 16% for CEA and from 5% to 22% for CAS. Perioperative stroke rate was the lowest in the POA method (1.1% CEA, 1.8% CAS) compared with two other methods without POA information (1.4% and 9.5% CEA and 2.4% and 16.4% CAS). Kappa indicated a poor (0.2) to fair (0.7) agreement between these approaches.

CONCLUSIONS

Administrative data has known limitations for assignment of symptom status and nonfatal perioperative outcomes. Given the uncertain timing of POA events as preoperative vs intraoperative and its apparent underestimation of the perioperative stroke rate, the use of administrative data even with the POA indicator for symptom status and non-fatal outcomes after CEA and CAS is hazardous.

Incorporating initial treatments improves performance of a mortality prediction model for patients with sepsis

PURPOSE

Mortality prediction models can be used to adjust for presenting severity of illness in observational studies of treatment effectiveness. We aimed to determine the incremental benefit of adding information about critical care services to a sepsis mortality prediction model.

METHODS

In a retrospective cohort of 166 931 eligible sepsis patients at 309 hospitals, we developed nested logistic regression models to predict mortality at the patient level. Our initial model included only demographic information. We then added progressively more detailed information such as comorbidities and initial treatments. We calculated each model's area under the receiver operating characteristic curve (AUROC) and also used a sheaf coefficient analysis to determine the relative effect of each additional group of variables.

RESULTS

Model discrimination increased as more detailed patient information was added. With demographics alone, the AUROC was 0.59; adding comorbidities increased the AUROC to 0.67. The final model, which took into account mixed (hierarchical) effects at the hospital level as well as initial treatments administered within the first two hospital days, resulted in an AUROC of 0.78. The standardized sheaf coefficient for the initial treatments was approximately 30% greater than that for demographics or infection source.

CONCLUSIONS

A sepsis disease risk score that incorporates information about the use of mechanical ventilation and vasopressors is superior to models that rely only on demographic information and comorbidities. Until administrative datasets include clinical information (such as vital signs and laboratory results), models such as this one could allow researchers to conduct observational studies of treatment effectiveness in sepsis patients.

Predicting healthcare utilization using a pharmacy-based metric with the WHO's Anatomic Therapeutic Chemical algorithm

BACKGROUND

Automated pharmacy claim data have been used for risk adjustment on health care utilization. However, most published pharmacy-based morbidity measures incorporate a coding algorithm that requires the medication data to be coded using the US National Drug Codes or the American Hospital Formulary Service drug codes, making studies conducted outside the US operationally cumbersome.

OBJECTIVE

This study aimed to verify that the pharmacy-based metric with the World Health Organization (WHO) Anatomical Therapeutic Chemical (ATC) algorithm can be used to explain the variations in health care utilization.

RESEARCH DESIGN

The Longitudinal Health Insurance Database of Taiwan's National Health Insurance enrollees was used in this study. We chose 2006 as the baseline year to predict the total cost, medication cost, and the number of outpatient visits in 2007. The pharmacy-based metric with 32 classes of chronic conditions was modified from a revised version of the Chronic Disease Score.

RESULTS

The ordinary least squares (OLS) model and log-transformed OLS model adjusted for the pharmacy-based metric had a better R in concurrently predicting total cost compared with the model adjusted for Deyo's Charlson Comorbidity Index and Elixhauser's Index. The pharmacy-based metric models also provided a superior performance in predicting medication cost and number of outpatient visits. For prospectively predicting health care utilization, the pharmacy-based metric models also performed better than the models adjusted by the diagnosis-based indices.

CONCLUSIONS

The pharmacy-based metric with the WHO ATC algorithm and the matching ATC codes were tested and found to be valid for explaining the variation in health care utilization.

Geographic variation in diagnosis frequency and risk of death among Medicare beneficiaries

CONTEXT

Because diagnosis is typically thought of as purely a patient attribute, it is considered a critical factor in risk-adjustment policies designed to reward efficient and high-quality care.

OBJECTIVE

To determine the association between frequency of diagnoses for chronic conditions in geographic areas and case-fatality rate among Medicare beneficiaries.

DESIGN, SETTING, AND PARTICIPANTS

Cross-sectional analysis of the mean number of 9 serious chronic conditions (cancer, chronic obstructive pulmonary disease, coronary artery disease, congestive heart failure, peripheral artery disease, severe liver disease, diabetes with end-organ disease, chronic renal failure, and dementia) diagnosed in 306 hospital referral regions (HRRs) in the United States; HRRs were divided into quintiles of diagnosis frequency. Participants were 5,153,877 fee-for-service Medicare beneficiaries in 2007.

MAIN OUTCOME MEASURES

Age/sex/race-adjusted case-fatality rates.

RESULTS

Diagnosis frequency ranged across HRRs from 0.58 chronic conditions in Grand Junction, Colorado, to 1.23 in Miami, Florida (mean, 0.90 [95% confidence interval {CI}, 0.89-0.91]; median, 0.87 [interquartile range, 0.80-0.96]). The number of conditions diagnosed was related to risk of death: among patients diagnosed with 0, 1, 2, and 3 conditions the case-fatality rate was 16, 45, 93, and 154 per 1000, respectively. As regional diagnosis frequency increased, however, the case fatality associated with a chronic condition became progressively less. Among patients diagnosed with 1 condition, the case-fatality rate decreased in a stepwise fashion across quintiles of diagnosis frequency, from 51 per 1000 in the lowest quintile to 38 per 1000 in the highest quintile (relative rate, 0.74 [95% CI, 0.72-0.76]). For patients diagnosed with 3 conditions, the corresponding case-fatality rates were 168 and 137 per 1000 (relative rate, 0.81 [95% CI, 0.79-0.84]).

CONCLUSION

Among fee-for-service Medicare beneficiaries, there is an inverse relationship between the regional frequency of diagnoses and the case-fatality rate for chronic conditions.

Development and validation of a disease-specific risk adjustment system using automated clinical data

OBJECTIVE

To develop and validate a disease-specific automated inpatient mortality risk adjustment system primarily using computerized numerical laboratory data and supplementing them with administrative data. To assess the values of additional manually abstracted data.

METHODS

Using 1,271,663 discharges in 2000-2001, we derived 39 disease-specific automated clinical models with demographics, laboratory findings on admission, ICD-9 principal diagnosis subgroups, and secondary diagnosis-based chronic conditions. We then added manually abstracted clinical data to the automated clinical models (manual clinical models). We compared model discrimination, calibration, and relative contribution of each group of variables. We validated these 39 models using 1,178,561 discharges in 2004-2005.

RESULTS

The overall mortality was 4.6 percent (n = 58,300) and 4.0 percent (n = 47,279) for derivation and validation cohorts, respectively. Common mortality predictors included age, albumin, blood urea nitrogen or creatinine, arterial pH, white blood counts, glucose, sodium, hemoglobin, and metastatic cancer. The average c-statistic for the automated clinical models was 0.83. Adding manually abstracted variables increased the average c-statistic to 0.85 with better calibration. Laboratory results displayed the highest relative contribution in predicting mortality.

CONCLUSIONS

A small number of numerical laboratory results and administrative data provided excellent risk adjustment for inpatient mortality for a wide range of clinical conditions.

Comparison of Rx-defined morbidity groups and diagnosis- based risk adjusters for predicting healthcare costs in Taiwan

Background

Medication claims are commonly used to calculate the risk adjustment for measuring healthcare cost. The Rx-defined Morbidity Groups (Rx-MG) which combine the use of medication to indicate morbidity have been incorporated into the Adjusted Clinical Groups (ACG) Case Mix System, developed by the Johns Hopkins University. This study aims to verify that the Rx-MG can be used for adjusting risk and for explaining the variations in the healthcare cost in Taiwan.

Methods

The Longitudinal Health Insurance Database 2005 (LHID2005) was used in this study. The year 2006 was chosen as the baseline to predict healthcare cost (medication and total cost) in 2007. The final sample size amounted to 793 239 (81%) enrolees, and excluded any cases with discontinued enrolment. Two different kinds of models were built to predict cost: the concurrent model and the prospective model. The predictors used in the predictive models included age, gender, Aggregated Diagnosis Groups (ADG, diagnosis- defined morbidity groups), and Rx-defined Morbidity Groups. Multivariate OLS regression was used in the cost prediction modelling.

Results

The concurrent model adjusted for Rx-defined Morbidity Groups for total cost, and controlled for age and gender had a better predictive R-square = 0.618, compared to the model adjusted for ADGs (R² = 0.411). The model combined with Rx-MGs and ADGs performed the best for concurrently predicting total cost (R² = 0.650). For prospectively predicting total cost, the model combined Rx-MGs and ADGs (R² = 0.382) performed better than the models adjusted by Rx-MGs (R² = 0.360) or ADGs (R² = 0.252) only. Similarly, the concurrent model adjusted for Rx-MGs predicting pharmacy cost had a better performance (R-square = 0.615), than the model adjusted for ADGs (R² = 0.431). The model combined with Rx-MGs and ADGs performed the best in concurrently as well as prospectively predicting pharmacy cost (R² = 0.638 and 0.505, respectively). The prospective models showed a remarkable improvement when adjusted by prior cost.

Conclusions

The medication-based Rx-Defined Morbidity Groups was useful in predicting pharmacy cost as well as total cost in Taiwan. Combining the information on medication and diagnosis as adjusters could arguably be the best method for explaining variations in healthcare cost.

The evolution of diagnosis-related groups (DRGs): from its beginnings in case-mix and resource use theory, to its implementation for payment and now for its current utilization for quality within and outside the hospital

Policymakers are searching for ways to control health care costs and improve quality. Diagnosis-related groups (DRGs) are by far the most important cost control and quality improvement tool that governments and private payers have implemented. This article reviews why DRGs have had this singular success both in the hospital sector and, over the past 10 years, in ambulatory and managed care settings. Last, the author reviews current trends in the development and implementation of tools that have the key ingredients of DRG success: categorical clinical model, separation of the clinical model from payment weights, separate payment adjustments for nonclinical factors, and outlier payments. Virtually all current tools used to manage health care costs and improve quality do not have these characteristics. This failure explains a key reason for the failure, for example, of the Medicare Advantage program to control health care costs. This article concludes with a discussion of future developments for DRG-type models outside the hospital sector.