Performance metrics for models designed to predict treatment effect

BACKGROUND

Measuring the performance of models that predict individualized treatment effect is challenging because the outcomes of two alternative treatments are inherently unobservable in one patient. The C-for-benefit was proposed to measure discriminative ability. However, measures of calibration and overall performance are still lacking. We aimed to propose metrics of calibration and overall performance for models predicting treatment effect in randomized clinical trials (RCTs).

METHODS

Similar to the previously proposed C-for-benefit, we defined observed pairwise treatment effect as the difference between outcomes in pairs of matched patients with different treatment assignment. We match each untreated patient with the nearest treated patient based on the Mahalanobis distance between patient characteristics. Then, we define the Eavg-for-benefit, E50-for-benefit, and E90-for-benefit as the average, median, and 90th quantile of the absolute distance between the predicted pairwise treatment effects and local-regression-smoothed observed pairwise treatment effects. Furthermore, we define the cross-entropy-for-benefit and Brier-for-benefit as the logarithmic and average squared distance between predicted and observed pairwise treatment effects. In a simulation study, the metric values of deliberately "perturbed models" were compared to those of the data-generating model, i.e., "optimal model". To illustrate these performance metrics, different modeling approaches for predicting treatment effect are applied to the data of the Diabetes Prevention Program: 1) a risk modelling approach with restricted cubic splines; 2) an effect modelling approach including penalized treatment interactions; and 3) the causal forest.

RESULTS

As desired, performance metric values of "perturbed models" were consistently worse than those of the "optimal model" (Eavg-for-benefit ≥ 0.043 versus 0.002, E50-for-benefit ≥ 0.032 versus 0.001, E90-for-benefit ≥ 0.084 versus 0.004, cross-entropy-for-benefit ≥ 0.765 versus 0.750, Brier-for-benefit ≥ 0.220 versus 0.218). Calibration, discriminative ability, and overall performance of three different models were similar in the case study. The proposed metrics were implemented in a publicly available R-package "HTEPredictionMetrics".

CONCLUSION

The proposed metrics are useful to assess the calibration and overall performance of models predicting treatment effect in RCTs.

Epidemiology of chronic inflammatory demyelinating polyradiculoneuropathy in The Netherlands

BACKGROUND AND AIMS

Chronic inflammatory demyelinating polyradiculoneuropathy (CIDP) is a rare but disabling disorder that often requires long-term immunomodulatory treatment. Background incidence rates and prevalence and risk factors for developing CIDP are still poorly defined. In the current study, we used a longitudinal population-based cohort study in The Netherlands to assess these rates and demographic factors and comorbidity associated with CIDP.

METHODS

We determined the incidence rate and prevalence of CIDP between 2008-2017 and the occurrence of potential risk factors in a retrospective Dutch cohort study using the Integrated Primary Care Information (IPCI) database. Cases were defined as CIDP if the diagnosis of CIDP was described in the electronic medical file.

RESULTS

In a source population of 928,030 persons with a contributing follow-up of 3,525,686 person-years, we identified 65 patients diagnosed with CIDP. The overall incidence rate was 0.68 per 100,000 person-years (95% CI 0.45-0.99). The overall prevalence was 7.00 per 100,000 individuals (95% CI 5.41-8.93). The overall incidence rate was higher in men compared to woman (IRR 3.00, 95% CI 1.27-7.11), and higher in elderly of 50 years or older compared to people <50 years of age (IRR 17 95% CI 4-73). Twenty percent of CIDP cases had DM and 9% a co-existing other autoimmune disease.

INTERPRETATION

These background rates are important to monitor changes in the frequency of CIDP following infectious disease outbreaks identify potential risk factors, and to estimate the social and economic burden of CIDP.

Instrumental variable analysis to estimate treatment effects: a simulation study showing potential benefits of conditioning on hospital

Background

Instrumental variable (IV) analysis holds the potential to estimate treatment effects from observational data. IV analysis potentially circumvents unmeasured confounding but makes a number of assumptions, such as that the IV shares no common cause with the outcome. When using treatment preference as an instrument, a common cause, such as a preference regarding related treatments, may exist. We aimed to explore the validity and precision of a variant of IV analysis where we additionally adjust for the provider: adjusted IV analysis.

Methods

A treatment effect on an ordinal outcome was simulated (beta − 0.5 in logistic regression) for 15.000 patients, based on a large data set (the IMPACT data, n = 8799) using different scenarios including measured and unmeasured confounders, and a common cause of IV and outcome. We compared estimated treatment effects with patient-level adjustment for confounders, IV with treatment preference as the instrument, and adjusted IV, with hospital added as a fixed effect in the regression models.

Results

The use of patient-level adjustment resulted in biased estimates for all the analyses that included unmeasured confounders, IV analysis was less confounded, but also less reliable. With correlation between treatment preference and hospital characteristics (a common cause) estimates were skewed for regular IV analysis, but not for adjusted IV analysis.

Conclusion

When using IV analysis for comparing hospitals, some limitations of regular IV analysis can be overcome by adjusting for a common cause.

Trial registration

We do not report the results of a health care intervention.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12874-022-01598-6.

Between-centre differences in care for in-hospital cardiac arrest: a prospective cohort study

BACKGROUND

Survival after in-hospital cardiac arrest is poor, but current literature shows substantial heterogeneity in reported survival rates. This study aims to evaluate care for patients suffering in-hospital cardiac arrest (IHCA) in the Netherlands by assessing between-hospital heterogeneity in outcomes and to explain this heterogeneity stemming from differences in case-mix or differences in quality of care.

METHODS

A prospective multicentre study was conducted comprising 14 centres. All IHCA patients were included. The adjusted variation in structure and process indicators of quality of care and outcomes (in-hospital mortality and cerebral performance category [CPC] scale) was assessed with mixed effects regression with centre as random intercept. Variation was quantified using the median odds ratio (MOR), representing the expected odds ratio for poor outcome between two randomly picked centres.

RESULTS

After excluding centres with less than 10 inclusions (2 centres), 701 patients were included of whom, 218 (32%) survived to hospital discharge. The unadjusted and case-mix adjusted MOR for mortality was 1.19 and 1.05, respectively. The unadjusted and adjusted MOR for CPC score was 1.24 and 1.19, respectively. In hospitals where personnel received cardiopulmonary resuscitation (CPR) training twice per year, 183 (64.7%) versus 290 (71.4%) patients died or were in a vegetative state, and 59 (20.8%) versus 68 (16.7%) patients showed full recovery (p < 0.001).

CONCLUSION

In the Netherlands, survival after IHCA is relatively high and between-centre differences in outcomes are small. The existing differences in survival are mainly attributable to differences in case-mix. Variation in neurological outcome is less attributable to case-mix.

Ordinal outcome analysis improves the detection of between-hospital differences in outcome

Background

There is a growing interest in assessment of the quality of hospital care, based on outcome measures. Many quality of care comparisons rely on binary outcomes, for example mortality rates. Due to low numbers, the observed differences in outcome are partly subject to chance. We aimed to quantify the gain in efficiency by ordinal instead of binary outcome analyses for hospital comparisons. We analyzed patients with traumatic brain injury (TBI) and stroke as examples.

Methods

We sampled patients from two trials. We simulated ordinal and dichotomous outcomes based on the modified Rankin Scale (stroke) and Glasgow Outcome Scale (TBI) in scenarios with and without true differences between hospitals in outcome. The potential efficiency gain of ordinal outcomes, analyzed with ordinal logistic regression, compared to dichotomous outcomes, analyzed with binary logistic regression was expressed as the possible reduction in sample size while keeping the same statistical power to detect outliers.

Results

In the IMPACT study (9578 patients in 265 hospitals, mean number of patients per hospital = 36), the analysis of the ordinal scale rather than the dichotomized scale (‘unfavorable outcome’), allowed for up to 32% less patients in the analysis without a loss of power. In the PRACTISE trial (1657 patients in 12 hospitals, mean number of patients per hospital = 138), ordinal analysis allowed for 13% less patients. Compared to mortality, ordinal outcome analyses allowed for up to 37 to 63% less patients.

Conclusions

Ordinal analyses provide the statistical power of substantially larger studies which have been analyzed with dichotomization of endpoints. We advise to exploit ordinal outcome measures for hospital comparisons, in order to increase efficiency in quality of care measurements.

Trial registration

We do not report the results of a health care intervention.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12874-020-01185-7.

Prediction admission in the older population in the Emergency Department: the CLEARED tool

BACKGROUND

Length of stay (LOS) in the Emergency Department (ED) is correlated with an extended in-hospital LOS and may even increase 30-day mortality. Older patients represent a growing population in the ED and they are especially at risk of adverse outcomes. Screening tools that adequately predict admission could help reduce waiting times in the ED and reduce time to treatment. We aimed to develop and validate a clinical prediction tool for admission, applicable to the aged patient population in the ED.

METHODS

Data from 7,606 ED visits of patients aged 70 years and older between 2012 and 2014 were used to develop the CLEARED tool. Model performance was assessed with discrimination using logistic regression and calibration. The model was internally validated by bootstrap resampling in Erasmus Medical Center and externally validated at two other hospitals, Medisch Spectrum Twente (MST) and Leiden University Medical Centre (LUMC).

RESULTS

CLEARED contains 10 predictors: body temperature, heart rate, diastolic blood pressure, systolic blood pressure, oxygen saturation, respiratory rate, referral status, the Manchester Triage System category, and the need for laboratory or radiology testing. The internally validated area under the curve (AUC) was 0.766 (95% CI [0.759;0.781]). External validation in MST showed an AUC of 0.797 and in LUMC, an AUC of 0.725.

CONCLUSIONS

The developed CLEARED tool reliably predicts admission in elderly patients visiting the ED. It is a promising prompt, although further research is needed to implement the tool and to investigate the benefits in terms of reduction of crowding and LOS in the ED.

Randomized trial of intravenous immunoglobulin maintenance treatment regimens in chronic inflammatory demyelinating polyradiculoneuropathy

Background and purpose

High peak serum immunoglobulin G (IgG) levels may not be needed for maintenance intravenous immunoglobulin (IVIg) treatment in chronic inflammatory demyelinating polyradiculoneuropathy (CIDP) and such high levels may cause side effects. More frequent lower dosing may lead to more stable IgG levels and higher trough levels, which might improve efficacy. The aim of this trial is to investigate whether high frequent low dosage IVIg treatment is more effective than low frequent high dosage IVIg treatment.

Methods

In this randomized placebo‐controlled crossover trial, we included patients with CIDP proven to be IVIg‐dependent and receiving an individually established stable dose and interval of IVIg maintenance treatment. In the control arm, patients received their individual IVIg dose and interval followed by a placebo infusion at half the interval. In the intervention arm, patients received half their individual dose at half the interval. After a wash‐out phase patients crossed over. The primary outcome measure was handgrip strength (assessed using a Martin Vigorimeter). Secondary outcome indicators were health‐related quality of life (36‐item Short‐Form Health Survey), disability (Inflammatory Rasch‐built Overall Disability Scale), fatigue (Rasch‐built Fatigue Severity Scale) and side effects.

Results

Twenty‐five patients were included and were treated at baseline with individually adjusted dosages of IVIg ranging from 20 to 80 g and intervals ranging from 14 to 35 days. Three participants did not complete the trial; the main analysis was therefore based on the 22 patients completing both treatment periods. There was no significant difference in handgrip strength change from baseline between the two treatment regimens (coefficient −2.71, 95% CI −5.4, 0.01). Furthermore, there were no significant differences in any of the secondary outcomes or side effects.

Conclusions

More frequent lower dosing does not further improve the efficacy of IVIg in stable IVIg‐dependent CIDP and does not result in fewer side effects.

Path From Clinical Research to Implementation: Endovascular Treatment of Ischemic Stroke in the Netherlands

Before 2015, endovascular treatment (EVT) for acute ischemic stroke was considered a promising treatment option. Based on limited evidence, it was performed in several dedicated stroke centers worldwide on selected patients. Since 2015, EVT for patients with intracranial large vessel occlusion has quickly been implemented as standard treatment in many countries worldwide, supported by the revised international guidelines based on solid evidence from multiple clinical trials. We describe the development in use of EVT in the Netherlands before, during, and after the pivotal EVT trials. We used data from all patients who were treated with EVT in the Netherlands from January 2002 until December 2018. We undertook a time-series analysis to examine trends in the use of EVT using Poisson regression analysis. Incidence rate ratios per year with 95% CIs were obtained to demonstrate the impact and implementation after the publication of the EVT trial results. We made regional observation plots, adjusted for stroke incidence, to assess the availability and use of the treatment in the country. In the buildup to the MR CLEAN (Multicenter Clinical Trial of Endovascular Treatment of Acute Ischemic Stroke in the Netherlands), a slow increase of EVT patients was observed, with 0.2% of all ischemic stroke patients receiving EVT. Before the trial results were formally announced, a statistically significant increase in EVT-treated patients per year was observed (incidence rate ratio, 1.72 [95% CI, 1.46-2.04]), and after the trial publication, an immediate steep increase was seen, followed by a more gradual increase (incidence rate ratio, 2.14 [95% CI, 1.77-2.59]). In 2018, the percentage of ischemic stroke patients receiving EVT increased to 5.8%. A well-developed infrastructure, a pragmatic approach toward the use of EVT in clinical practice, in combination with a strict adherence by the regulatory authorities to national evidence-based guidelines has led to successful implementation of EVT in the Netherlands. Ongoing efforts are directed at further increasing the proportion of stroke patients with EVT in all regions of the country.

Low Socioeconomic Status Is Associated with Worse Outcomes After Curative Surgery for Colorectal Cancer: Results from a Large, Multicenter Study

BACKGROUND

Socioeconomic status (SES) has been associated with early mortality in cancer patients. However, the association between SES and outcome in colorectal cancer patients is largely unknown. The aim of this study was to investigate whether SES is associated with short- and long-term outcome in patients undergoing curative surgery for colorectal cancer.

METHODS

Patients who underwent curative surgery in the region of Rotterdam for stage I-III colorectal cancer between January 2007 and July 2014 were included. Gross household income and survival status were obtained from a national registry provided by Statistics Netherlands Centraal Bureau voor de Statistiek. Patients were assigned percentiles according to the national income distribution. Logistic regression and Cox proportional hazard regression were performed to assess the association of SES with 30-day postoperative complications, overall survival and cancer-specific survival, adjusted for known prognosticators.

RESULTS

For 965 of the 975 eligible patients (99%), gross household income could be retrieved. Patients with a lower SES more often had diabetes, more often underwent an open surgical procedure, and had more comorbidities. In addition, patients with a lower SES were less likely to receive (neo) adjuvant treatment. Lower SES was independently associated with an increased risk of postoperative complications (Odds ratio per percent increase 0.99, 95%CI 0.99-0.998, p = 0.004) and lower cancer-specific mortality (Hazard ratio per percent increase 0.99, 95%CI 0.98-0.99, p = 0.009).

CONCLUSION

This study shows that lower SES is associated with increased risk of postoperative complications, and poor cancer-specific survival in patients undergoing surgery for stage I-III colorectal cancer after correcting for known prognosticators.

Variation in the practice of tracheal intubation in Europe after traumatic brain injury: a prospective cohort study

Traumatic brain injury patients frequently undergo tracheal intubation. We aimed to assess current intubation practice in Europe and identify variation in practice. We analysed data from patients with traumatic brain injury included in the prospective cohort study collaborative European neurotrauma effectiveness research in traumatic brain injury (CENTER‐TBI) in 45 centres in 16 European countries. We included patients who were transported to hospital by emergency medical services. We used mixed‐effects multinomial regression to quantify the effects on pre‐hospital or in‐hospital tracheal intubation of the following: patient characteristics; injury characteristics; centre; and trauma system characteristics. A total of 3843 patients were included. Of these, 1322 (34%) had their tracheas intubated; 839 (22%) pre‐hospital and 483 (13%) in‐hospital. The fit of the model with only patient characteristics predicting intubation was good (Nagelkerke R2 64%). The probability of tracheal intubation increased with the following: younger age; lower pre‐hospital or emergency department GCS; higher abbreviated injury scale scores (head and neck, thorax and chest, face or abdomen abbreviated injury score); and one or more unreactive pupils. The adjusted median odds ratio for intubation between two randomly chosen centres was 3.1 (95%CI 2.1–4.3) for pre‐hospital intubation, and 2.7 (95%CI 1.9–3.5) for in‐hospital intubation. Furthermore, the presence of an anaesthetist was independently associated with more pre‐hospital intubation (OR 2.9, 95%CI 1.3–6.6), in contrast to the presence of ambulance personnel who are allowed to intubate (OR 0.5, 95%CI 0.3–0.8). In conclusion, patient and injury characteristics are key drivers of tracheal intubation. Between‐centre differences were also substantial. Further studies are needed to improve the evidence base supporting recommendations for tracheal intubation.

Trauma models to identify major trauma and mortality in the prehospital setting

BACKGROUND

Patients with major trauma might benefit from treatment in a trauma centre, but early identification of major trauma (Injury Severity Score (ISS) over 15) remains difficult. The aim of this study was to undertake an external validation of existing prognostic models for injured patients to assess their ability to predict mortality and major trauma in the prehospital setting.

METHODS

Prognostic models were identified through a systematic literature search up to October 2017. Injured patients transported by Emergency Medical Services to an English hospital from the Trauma Audit and Research Network between 2013 and 2016 were included. Outcome measures were major trauma (ISS over 15) and in-hospital mortality. The performance of the models was assessed in terms of discrimination (concordance index, C-statistic) and net benefit to assess the clinical usefulness.

RESULTS

A total of 154 476 patients were included to validate six previously proposed prediction models. Discriminative ability ranged from a C-statistic value of 0·602 (95 per cent c.i. 0·596 to 0·608) for the Mechanism, Glasgow Coma Scale, Age and Arterial Pressure model to 0·793 (0·789 to 0·797) for the modified Rapid Emergency Medicine Score (mREMS) in predicting in-hospital mortality (11 882 patients). Major trauma was identified in 52 818 patients, with discrimination from a C-statistic value of 0·589 (0·586 to 0·592) for mREMS to 0·735 (0·733 to 0·737) for the Kampala Trauma Score in predicting major trauma. None of the prediction models met acceptable undertriage and overtriage rates.

CONCLUSION

Currently available prehospital trauma models perform reasonably in predicting in-hospital mortality, but are inadequate in identifying patients with major trauma. Future research should focus on which patients would benefit from treatment in a major trauma centre.

Prediction models for mortality in adult patients visiting the Emergency Department: a systematic review

We provide a systematic overview of literature on prediction models for mortality in the Emergency Department (ED). We searched various databases for observational studies in the ED or similar setting describing prediction models for short-term mortality (up to 30 days or in-hospital mortality) in a non-trauma population. We used the CHARMS-checklist for quality assessment. We found a total of 14.768 articles and included 17 articles, describing 22 models. Model performance ranged from AUC 0.63-0.93. Most articles had a moderate risk of bias in one or more domains. The full model and PARIS model performed best, but are not yet ready for implementation. There is a need for validation studies to compare multiple prediction models and to evaluate their accuracy.

Patient-Reported Outcome Measures May Add Value in Breast Cancer Surgery

PURPOSE

Considering the comparable prognosis in early-stage breast cancer after breast-conserving therapy (BCT) and mastectomy, quality of life should be a focus in treatment decision(s). We retrospectively collected PROs and analyzed differences per type of surgery delivered. We aimed to obtain reference values helpful in shared decision-making.

PATIENTS AND METHODS

pTis-T3N0-3M0 patients operated between January 2005 and September 2016 were eligible if: (1) no chemotherapy was administered < 6 months prior to enrolment, and (2) identical surgeries were performed in case of bilateral surgery. After consent, EQ-5D-5L, EORTC-QLQ-C30/BR23, and BREAST-Q were administered. PROs were evaluated per baseline characteristics using multivariable linear regression models. Outcomes were compared for different surgeries as well as for primary (PBC) and second primary or recurrent (SBC) breast cancer patients using analyses of variance (ANOVAs).

RESULTS

The response rate was 68%. PROs in 612 PBC patients were comparable to those in 152 SBC patients. Multivariable analyses showed increasing age to be associated with lower "physical functioning" [β - 0.259, p < 0.001] and "sexual functioning" [β - 0.427, p < 0.001], and increasing time since surgery with less "fatigue" [β - 1.083, p < 0.001]. Mastectomy [β - 13.596, p = 0.003] and implant reconstruction [β - 13.040, p = 0.007] were associated with lower "satisfaction with breast" scores than BCT. Radiation therapy was associated with lower satisfaction scores than absence of radiotherapy.

DISCUSSION

PRO scores were associated with age, time since surgery, type of surgery, and radiation therapy in breast cancer patients. The scores serve as a reference value for different types of surgery in the study population and enable prospective use of PROs in shared decision-making.

Value of Thrombus CT Characteristics in Patients with Acute Ischemic Stroke

BACKGROUND AND PURPOSE

Thrombus CT characteristics might be useful for patient selection for intra-arterial treatment. Our objective was to study the association of thrombus CT characteristics with outcome and treatment effect in patients with acute ischemic stroke.

MATERIALS AND METHODS

We included 199 patients for whom thin-section NCCT and CTA within 30 minutes from each other were available in the Multicenter Randomized Clinical Trial of Endovascular Treatment for Acute ischemic stroke in the Netherlands (MR CLEAN) study. We assessed the following thrombus characteristics: location, distance from ICA terminus to thrombus, length, volume, absolute and relative density on NCCT, and perviousness. Associations of thrombus characteristics with outcome were estimated with univariable and multivariable ordinal logistic regression as an OR for a shift toward better outcome on the mRS. Interaction terms were used to investigate treatment-effect modification by thrombus characteristics.

RESULTS

In univariate analysis, only the distance from the ICA terminus to the thrombus, length of >8 mm, and perviousness were associated with functional outcome. Relative thrombus density on CTA was independently associated with functional outcome with an adjusted common OR of 1.21 per 10% (95% CI, 1.02-1.43; P = .029). There was no treatment-effect modification by any of the thrombus CT characteristics.

CONCLUSIONS

In our study on patients with large-vessel occlusion of the anterior circulation, CT thrombus characteristics appear useful for predicting functional outcome. However, in our study cohort, the effect of intra-arterial treatment was independent of the thrombus CT characteristics. Therefore, no arguments were provided to select patients for intra-arterial treatment using thrombus CT characteristics.

Validation of a base deficit-based trauma prediction model and comparison with TRISS and ASCOT

BACKGROUND

Base deficit provides a more objective indicator of physiological stress following injury as compared with vital signs constituting the revised trauma score (RTS). We have previously developed a base deficit-based trauma survival prediction model [base deficit and injury severity score model (BISS)], in which RTS was replaced by base deficit as a measurement of physiological imbalance.

PURPOSE

To externally validate BISS in a large cohort of trauma patients and to compare its performance with established trauma survival prediction models including trauma and injury severity score (TRISS) and a severity characterization of trauma (ASCOT). Moreover, we examined whether the predictive accuracy of BISS model could be improved by replacement of injury severity score (ISS) by new injury severity score (NISS) in the BISS model (BNISS).

METHODS

In this retrospective, observational study, clinical data of 3737 trauma patients (age ≥15 years) admitted consecutively from 2003 to 2007 were obtained from a prospective trauma registry to calculate BISS, TRISS, and ASCOT models. The models were evaluated in terms of discrimination [area under curve (AUC)] and calibration.

RESULTS

The in-hospital mortality rate was 8.1 %. The discriminative performance of BISS to predict survival was similar to that of TRISS and ASCOT [AUCs of 0.883, 95 % confidence interval (CI) 0.865-0.901 for BISS, 0.902, 95 % CI 0.858-0.946 for TRISS and 0.864, 95 % CI 0.816-0.913 for ASCOT]. Calibration tended to be optimistic in all three models. The updated BNISS had an AUC of 0.918 indicating that substitution of ISS with NISS improved model performance.

CONCLUSIONS

The BISS model, a base deficit-based trauma model for survival prediction, showed equivalent performance as compared with that of TRISS and ASCOT and may offer a more simplified calculation method and a more objective assessment. Calibration of BISS model was, however, less good than that of other models. Replacing ISS by NISS can considerably improve model accuracy, but further confirmation is needed.

Comparing colon cancer outcomes: The impact of low hospital case volume and case-mix adjustment

OBJECTIVE

When comparing performance across hospitals it is essential to consider the noise caused by low hospital case volume and to perform adequate case-mix adjustment. We aimed to quantify the role of noise and case-mix adjustment on standardized postoperative mortality and anastomotic leakage (AL) rates.

METHODS

We studied 13,120 patients who underwent colon cancer resection in 85 Dutch hospitals. We addressed differences between hospitals in postoperative mortality and AL, using fixed (ignoring noise) and random effects (incorporating noise) logistic regression models with general and additional, disease specific, case-mix adjustment.

RESULTS

Adding disease specific variables improved the performance of the case-mix adjustment models for postoperative mortality (c-statistic increased from 0.77 to 0.81). The overall variation in standardized mortality ratios was similar, but some individual hospitals changed considerably. For the standardized AL rates the performance of the adjustment models was poor (c-statistic 0.59 and 0.60) and overall variation was small. Most of the observed variation between hospitals was actually noise.

CONCLUSION

Noise had a larger effect on hospital performance than extended case-mix adjustment, although some individual hospital outcome rates were affected by more detailed case-mix adjustment. To compare outcomes between hospitals it is crucial to consider noise due to low hospital case volume with a random effects model.

Including post-discharge mortality in calculation of hospital standardised mortality ratios: retrospective analysis of hospital episode statistics

OBJECTIVES

To assess the consequences of applying different mortality timeframes on standardised mortality ratios of individual hospitals and, secondarily, to evaluate the association between in-hospital standardised mortality ratios and early post-discharge mortality rate, length of hospital stay, and transfer rate.

DESIGN

Retrospective analysis of routinely collected hospital data to compare observed deaths in 50 diagnostic categories with deaths predicted by a case mix adjustment method.

SETTING

60 Dutch hospitals.

PARTICIPANTS

1 228 815 patients discharged in the period 2008 to 2010.

MAIN OUTCOME MEASURES

In-hospital standardised mortality ratio, 30 days post-admission standardised mortality ratio, and 30 days post-discharge standardised mortality ratio.

RESULTS

Compared with the in-hospital standardised mortality ratio, 33% of the hospitals were categorised differently with the 30 days post-admission standardised mortality ratio and 22% were categorised differently with the 30 days post-discharge standardised mortality ratio. A positive association was found between in-hospital standardised mortality ratio and length of hospital stay (Pearson correlation coefficient 0.33; P=0.01), and an inverse association was found between in-hospital standardised mortality ratio and early post-discharge mortality (Pearson correlation coefficient -0.37; P=0.004).

CONCLUSIONS

Applying different mortality timeframes resulted in differences in standardised mortality ratios and differences in judgment regarding the performance of individual hospitals. Furthermore, associations between in-hospital standardised mortality rates, length of stay, and early post-discharge mortality rates were found. Combining these findings suggests that standardised mortality ratios based on in-hospital mortality are subject to so-called "discharge bias." Hence, early post-discharge mortality should be included in the calculation of standardised mortality ratios.

Use of surgical-site infection rates to rank hospital performance across several types of surgery

BACKGROUND

Comparing and ranking hospitals based on health outcomes is becoming increasingly popular, although case-mix differences between hospitals and random variation are known to distort interpretation. The aim of this study was to explore whether surgical-site infection (SSI) rates are suitable for comparing hospitals, taking into account case-mix differences and random variation.

METHODS

Data from the national surveillance network in the Netherlands, on the eight most frequently registered types of surgery for the year 2009, were used to calculate SSI rates. The variation in SSI rate between hospitals was estimated with multivariable fixed- and random-effects logistic regression models to account for random variation and case mix. 'Rankability' (as the reliability of ranking) of the SSI rates was calculated by relating within-hospital variation to between-hospital variation.

RESULTS

Thirty-four hospitals reported on 13 629 patients, with overall SSI rates per surgical procedure varying between 0 and 15·1 per cent. Statistically significant differences in SSI rate between hospitals were found for colonic resection, caesarean section and for all operations combined. Rankability was 80 per cent for colonic resection but 0 per cent for caesarean section. Rankability was 8 per cent in all operations combined, as the differences in SSI rates were explained mainly by case mix.

CONCLUSION

When comparing SSI rates in all operations, differences between hospitals were explained by case mix. For individual types of surgery, case mix varied less between hospitals, and differences were explained largely by random variation. Although SSI rates may be used for monitoring quality improvement within hospitals, they should not be used for ranking hospitals.

Early recognition of poor prognosis in Guillain-Barre syndrome

BACKGROUND

Guillain-Barré syndrome (GBS) has a highly diverse clinical course and outcome, yet patients are treated with a standard therapy. Patients with poor prognosis may benefit from additional treatment, provided they can be identified early, when nerve degeneration is potentially reversible and treatment is most effective. We developed a clinical prognostic model for early prediction of outcome in GBS, applicable for clinical practice and future therapeutic trials.

METHODS

Data collected prospectively from a derivation cohort of 397 patients with GBS were used to identify risk factors of being unable to walk at 4 weeks, 3 months, and 6 months. Potential predictors of poor outcome (unable to walk unaided) were considered in univariable and multivariable logistic regression models. The clinical model was based on the multivariable logistic regression coefficients of selected predictors and externally validated in an independent cohort of 158 patients with GBS.

RESULTS

High age, preceding diarrhea, and low Medical Research Council sumscore at hospital admission and at 1 week were independently associated with being unable to walk at 4 weeks, 3 months, and 6 months (all p 0.05-0.001). The model can be used at hospital admission and at day 7 of admission, the latter having a better predictive ability for the 3 endpoints; the area under the receiver operating characteristic curve (AUC) is 0.84-0.87 and at admission the AUC is 0.73-0.77. The model proved to be valid in the validation cohort.

CONCLUSIONS

A clinical prediction model applicable early in the course of disease accurately predicts the first 6 months outcome in GBS.

Statin treatment after a recent TIA or stroke: is effectiveness shown in randomized clinical trials also observed in everyday clinical practice?

AIM AND BACKGROUND

The benefit of statin treatment in patients with a previous ischemic stroke or transient ischemic attack (TIA) has been demonstrated in randomized clinical trials (RCT). However, the effectiveness in everyday clinical practice may be decreased because of a different patient population and less controlled setting. We aim to describe statin use in an unselected cohort of patients, identify factors related to statin use and test whether the effect of statins on recurrent vascular events and mortality observed in RCTs is also observed in everyday clinical practice.

METHODS

In 10 centers in the Netherlands, patients admitted to the hospital or visiting the outpatient clinic with a recent TIA or ischemic stroke were prospectively and consecutively enrolled between October 2002 and May 2003. Statin use was determined at discharge and during follow-up. We used logistic regression models to estimate the effect of statins on the occurrence of vascular events (stroke or myocardial infarction) and mortality within 3 years. We adjusted for confounders with a propensity score that relates patient characteristics to the probability of using statins.

RESULTS

Of the 751 patients in the study, 252 (34%) experienced a vascular event within 3 years. Age, elevated cholesterol levels and other cardiovascular risk factors were associated with statin use at discharge. After 3 years, 109 of 280 (39%) of the users at discharge had stopped using statins. Propensity score adjusted analyses showed a beneficial effect of statins on the occurrence of the primary outcome (odds ratio 0.8, 95% CI: 0.6-1.2).

CONCLUSION

In our study, we found poor treatment adherence to statins. Nevertheless, after adjustment for the differences between statin users and non-statin users, the observed beneficial effect of statins on the occurrence of vascular events within 3 years, although not statistically significant, is compatible with the effect observed in clinical trials.

Process of care partly explains the variation in mortality between hospitals after peripheral vascular surgery

OBJECTIVES

The aim of this study is to investigate whether variation in mortality at hospital level reflects differences in quality of care of peripheral vascular surgery patients.

DESIGN

Observational study.

MATERIALS

In 11 hospitals in the Netherlands, 711 consecutive vascular surgery patients were enrolled.

METHODS

Multilevel logistic regression models were used to relate patient characteristics, structure and process of care to mortality at 1 year. The models were constructed by consecutively adding age, sex and Lee index, then remaining risk factors, followed by structural measures for quality of care and finally, selected process of care parameters.

RESULTS

Total 1-year mortality was 11%, ranging from 6% to 26% in different hospitals. Large differences in patient characteristics and quality indicators were observed between hospitals (e.g., age>70 years: 28-58%; beta-blocker therapy: 39-87%). Adjusted analyses showed that a large part of variation in mortality was explained by age, sex and the Lee index (Akaike's information criterion (AIC)=59, p<0.001). Another substantial part of the variation was explained by process of care (AIC=5, p=0.001).

CONCLUSIONS

Differences between hospitals exist in patient characteristics, structure of care, process of care and mortality. Even after adjusting for the patient population at risk, a substantial part of the variation in mortality can be explained by differences in process measures of quality of care.

Comparing and ranking hospitals based on outcome: results from The Netherlands Stroke Survey

BACKGROUND

Measuring quality of care and ranking hospitals with outcome measures poses two major methodological challenges: case-mix adjustment and variation that exists by chance.

AIM

To compare methods for comparing and ranking hospitals that considers these.

METHODS

The Netherlands Stroke Survey was conducted in 10 hospitals in the Netherlands, between October 2002 and May 2003, with prospective and consecutive enrollment of patients with acute brain ischaemia. Poor outcome was defined as death or disability after 1 year (modified Rankin scale of > or =3). We calculated fixed and random hospital effects on poor outcome, unadjusted and adjusted for patient characteristics. We compared the hospitals using the expected rank, a novel statistical measure incorporating the magnitude and the uncertainty of differences in outcome.

RESULTS

At 1 year after stroke, 268 of the total 505 patients (53%) had a poor outcome. There were substantial differences in outcome between hospitals in unadjusted analysis (chi(2) = 48, 9 df, P < 0.0001). Adjustment for 12 confounders led to halving of the chi(2) (chi(2) = 24). The same pattern was observed in random effects analysis. Estimated performance of individual hospitals changed considerably between unadjusted and adjusted analysis. Further changes were seen with random effect estimation, especially for smaller hospitals. Ordering by expected rank led to shrinkage of the original ranks of 1-10 towards the median rank of 5.5 and to a different order of the hospitals, compared to ranking based on fixed effects.

CONCLUSION

In comparing and ranking hospitals, case-mix-adjusted random effect estimates and the expected ranks are more robust alternatives to traditional fixed effect estimates and simple rankings.

New approaches to increase statistical power in TBI trials: insights from the IMPACT study

INTRODUCTION

None of the multi-centre phase III randomized controlled trials (RCTs) performed in TBI have convincingly demonstrated efficacy. Problems in clinical trial design and analysis may have contributed to these failures. Clinical trials in the TBI population pose several complicated methodological challenges, related especially to the heterogeneity of the population. In this paper we examine the issue of heterogeneity within the IMPACT (International Mission on Prognosis and Clinical Trial design in TBI) database and investigate the application of conventional and innovative methods for the statistical analysis of trials in TBI.

METHODS AND RESULTS

Simulation studies in the IMPACT database (N = 9205) showed substantial gains in efficiency with covariate adjustment. Adjusting for 7 important predictors yielded up to a 28% potential reduction in trial size. Ongoing analyses on the potential benefit of ordinal analysis, such as proportional odds and sliding dichotomy, gave promising results with even larger potential reductions in trial size.

CONCLUSION

The statistical power of RCTs in TBI can be considerably increased by applying covariate adjustment and by ordinal analysis methods of the GOS. These methods need to be considered for optimizing future TBI trials.

[Differences between hospitals in outcome after a stroke are only partially explained by differences in the quality of care]

OBJECTIVE

To determine the extent to which the outcome of stroke patients stroke is correlated with patient characteristics and care process parameters, and to determine whether outcome measures can be used to measure the quality of hospital care provided for these patients.

DESIGN

Descriptive cohort study.

METHODS

At 10 hospitals in the Netherlands, in the period October 2002-April 2003, patients with acute stroke were included in the study. Poor outcome was defined as dead or disabled at 1 year (a score on the modified Rankin scale > or = 3). Quality of the care was assessed by relating diagnostic, therapeutic and preventive procedures to indication. Multiple logistic regression models were used to compare observed numbers of patients with a poor outcome with expected numbers per hospital, after adjustment for patient characteristics and quality of care parameters.

RESULTS

In total, 579 patients were included in the study, of which 271 (47%) were dead or disabled at 1 year. Poor outcome varied across the hospitals from 29 to 78%. The mean age was 70 years. There were large differences between hospitals with respect to patient characteristics and quality of care. Most of the differences in outcome between hospitals were explained by the differences in patient characteristics (Akaike's information criterion (AIC) = 134). Quality of care parameters explained just a small additional part of the variation in patient outcome (AIC = 5.5).

CONCLUSIONS

Large differences between Dutch hospitals in the patient outcome after stroke could mostly be explained by differences in patient characteristics. Only a small part of the hospital variation in patient outcome was related to differences in quality of care. Therefore, outcome indicators cannot be regarded as valid performance indicators for care following a stroke.

Variation between hospitals in patient outcome after stroke is only partly explained by differences in quality of care: results from the Netherlands Stroke Survey

BACKGROUND AND PURPOSE

Patient outcome is often used as an indicator of quality of hospital care. The aim of this study is to investigate whether there is a straightforward relationship between quality of care and outcome, and whether outcome measures could be used to assess quality of care after stroke.

METHODS

In 10 centres in The Netherlands, 579 patients with acute stroke were prospectively and consecutively enrolled. Poor outcome was defined as a score on the modified Rankin scale >or=3 at 1 year. Quality of care was assessed by relating diagnostic, therapeutic and preventive procedures to indication. Multiple logistic regression models were used to compare observed proportions of patients with poor outcome with expected proportions, after adjustment for patient characteristics and quality of care parameters.

RESULTS

A total of 271 (47%) patients were dead or disabled at 1 year. Poor outcome varied across the centres from 29% to 78%. Large differences between centres were also observed in clinical characteristics, prognostic factors and quality of care. For example, between hospital quartiles based on outcome, age >or=70 years varied from 50% to 65%, presence of vascular risk factors from 88% to 96%, intravenous fluids when indicated from 35% to 81%, and antihypertensive therapy when indicated from 60% to 85%. The largest part of variation in patient outcome between centres was explained by differences in patient characteristics (Akaike's Information Criterion (AIC) = 134.0). Quality of care parameters explained a small part of the variation in patient outcome (AIC = 5.5).

CONCLUSIONS

Patient outcome after stroke varies largely between centres and is, for a substantial part, explained by differences in patient characteristics at time of hospital admission. Only a small part of the hospital variation in patient outcome is related to differences in quality of care. Unadjusted proportions of poor outcome after stroke are not valid as indicators of quality of care.