A multi-state model for evolution of intensive care unit patients: prediction of nosocomial infections and deaths

Efficacy of Random Forest Models in Predicting Multidrug-Resistant Gram-Negative Bacterial Nosocomial Infections Compared to Traditional Logistic Regression Models

This study evaluates whether random forest (RF) models are as effective as traditional Logistic Regression (LR) models in predicting multidrug-resistant Gram-negative bacterial nosocomial infections. Data were collected from 541 patients with hospital-acquired Gram-negative bacterial infections at two tertiary-level hospitals in Urumqi, Xinjiang, China, from August 2022 to November 2023. Relevant literature informed the selection of significant predictors based on patients' pre-infection clinical information and medication history. The data were split into a training set of 379 cases and a validation set of 162 cases, adhering to a 7:3 ratio. Both RF and LR models were developed using the training set and subsequently evaluated on the validation set. The LR model achieved an accuracy of 84.57%, sensitivity of 82.89%, specificity of 80.10%, positive predictive value of 84%, negative predictive value of 85.06%, and a Yoden index of 0.69. In contrast, the RF model demonstrated superior performance with an accuracy of 89.51%, sensitivity of 90.79%, specificity of 88.37%, positive predictive value of 87.34%, negative predictive value of 91.57%, and a Yoden index of 0.79. Receiver operating characteristic curve analysis revealed an area under the curve of 0.91 for the LR model and 0.94 for the RF model. These findings indicate that the RF model surpasses the LR model in specificity, sensitivity, and accuracy in predicting hospital-acquired multidrug-resistant Gram-negative infections, showcasing its greater potential for clinical application.

Healthcare-associated infections in patients with severe COVID-19 supported with extracorporeal membrane oxygenation: a nationwide cohort study

BACKGROUND

Both critically ill patients with coronavirus disease 2019 (COVID-19) and patients receiving extracorporeal membrane oxygenation (ECMO) support exhibit a high incidence of healthcare-associated infections (HAI). However, data on incidence, microbiology, resistance patterns, and the impact of HAI on outcomes in patients receiving ECMO for severe COVID-19 remain limited. We aimed to report HAI incidence and microbiology in patients receiving ECMO for severe COVID-19 and to evaluate the impact of ECMO-associated infections (ECMO-AI) on in-hospital mortality.

METHODS

For this study, we analyzed data from 701 patients included in the ECMOSARS registry which included COVID-19 patients supported by ECMO in France.

RESULTS

Among 602 analyzed patients for whom HAI and hospital mortality data were available, 214 (36%) had ECMO-AI, resulting in an incidence rate of 27 ECMO-AI per 1000 ECMO days at risk. Of these, 154 patients had bloodstream infection (BSI) and 117 patients had ventilator-associated pneumonia (VAP). The responsible microorganisms were Enterobacteriaceae (34% for BSI and 48% for VAP), Enterococcus species (25% and 6%, respectively) and non-fermenting Gram-negative bacilli (13% and 20%, respectively). Fungal infections were also observed (10% for BSI and 3% for VAP), as were multidrug-resistant organisms (21% and 15%, respectively). Using a Cox multistate model, ECMO-AI were not found associated with hospital death (HR = 1.00 95% CI [0.79-1.26], p = 0.986).

CONCLUSIONS

In a nationwide cohort of COVID-19 patients receiving ECMO support, we observed a high incidence of ECMO-AI. ECMO-AI were not found associated with hospital death. Trial registration number NCT04397588 (May 21, 2020).

The polytomous discrimination index for prediction involving multistate processes under intermittent observation

With the increasing importance of predictive modeling in health research comes the need for methods to rigorously assess predictive accuracy. We consider the problem of evaluating the accuracy of predictive models for nominal outcomes when outcome data are coarsened at random. We first consider the problem in the context of a multinomial response modeled by polytomous logistic regression. Attention is then directed to the motivating setting in which class membership corresponds to the state occupied in a multistate disease process at a time horizon of interest. Here, class (state) membership may be unknown at the time horizon since disease processes are under intermittent observation. We propose a novel extension to the polytomous discrimination index to address this and evaluate the predictive accuracy of an intensity-based model in the context of a study involving patients with arthritis from a registry at the University of Toronto Centre for Prognosis Studies in Rheumatic Diseases.

Mortality due to hospital-acquired infection after cardiac surgery

PURPOSE

Hospital-acquired infections have been associated with significant morbidity and mortality in critically ill surgical patients. However, little is known about mortality due to hospital-acquired infections in cardiac surgery.

METHODS

We conducted a retrospective analysis of prospectively collected data from the cardiac surgery unit of a university hospital. All patients who underwent cardiac surgery over a 7-year period were included. Patients with hospital-acquired infections were matched 1:1 with patients with nonhospital-acquired infections based on risk factors for hospital-acquired infections and death after cardiac surgery using propensity score matching. We performed a competitive risk analysis to study the mortality fraction due to hospital-acquired infections.

RESULTS

Of 8853 patients who underwent cardiac surgery, 370 (4.2%) developed 500 postoperative infections (incidence density rate 4.2 hospital-acquired infections per 1000 patient-days). Crude hospital mortality was significantly higher in patients with hospital-acquired infections than in matched patients who did not develop hospital-acquired infections, 15.4% and 5.7%, respectively (P < .001). The in-hospital mortality fraction due to hospital-acquired infections in our cohort was 17.1% (12.3%-22.8%). Pseudomonas aeruginosa infection (hazard ratio, 2.09; 95% confidence interval, 1.23-3.49; P = .005), bloodstream infection (hazard ratio, 2.08; 95% confidence interval, 1.19-3.63; P = .010), and pneumonia (hazard ratio, 1.68; 95% confidence interval, 1.02-2.77; P = .04) were each independently associated with increased hospital mortality.

CONCLUSIONS

Although hospital-acquired infections are relatively uncommon after cardiac surgery (4.2%), these infections have a major impact on postoperative mortality (attributable mortality fraction, 17.1%).

Attributable mortality of ventilator-associated pneumonia

PURPOSE OF REVIEW

To critically discuss the attributable mortality of ventilator-associated pneumonia (VAP) and potential sources of variation.

RECENT FINDINGS

The review will cover the available estimates (0-50%). It will also explore the source of variation because of definition of VAP (being lower if inaccurate), case-mix issues (being lower for trauma patients), the severity of underlying illnesses (being maximal when the severity of underlying illness is intermediate), and on the characteristics and the severity of the VAP episode. Another important source of variation is the use of poorly appropriate statistical models (estimates biased by lead time bias and competing events). New extensions of survival models which take into account the time dependence of VAP occurrence and competing risks allow less biased estimation as compared with traditional models.

SUMMARY

Attributable mortality of VAP is about 6%. Accurate diagnostic methods are key to properly estimating it. Traditional statistical models should no longer be used to estimate it. Prevention efforts targeted on patients with intermediate severity may result in the most important outcome benefits.

Prediction of long-term kidney failure in renal transplant with chronic allograft dysfunction using stage-specific hazard rates

OBJECTIVES

The process of kidney failure in renal transplant recipients with chronic allograft dysfunction is characterized by a progressive decline in glomerular filtration rate over time that it is determined by the 5-stage model. This study used stage-based statistical survival analysis to predict graft survival in renal transplant recipients with chronic allograft dysfunction.

MATERIALS AND METHODS

In a single-center, retrospective study, 214 renal transplant recipients with chronic allograft dysfunction were investigated at a university hospital in Iran from 1997 to 2005. At each patient visit, kidney function was assessed using glomerular filtration rate and stage of disease.

RESULTS

The estimated stage-specific hazard rates of disease progression are stage one, 453.936; stage two, 485.040; stage three, 545.808; and stage four; 649.488 per 1000 person-years. The estimated mean times in each stage were as follows: kidney damage with normal or increased glomerular filtration rate, 26.43 months; kidney damage with mildly decreased glomerular filtration rate, 24.74 months; moderate kidney disease, 21.98 months; and severe kidney disease; 18.48 months. These estimates yield a mean time from stage 1 to kidney failure of 91.63 months. The probability of graft survival was predicted using estimated stage-specific hazard rates. The 18th, 58th, 118th, and 155th months death-censored graft survival probabilities were 0.99, 0.75, 0.25, and 0.10.

CONCLUSIONS

In this method of survival analysis, we can determine a statistical model according to a real clinical model in renal transplant recipients with chronic allograft dysfunction. It enables us to determine the stage-specific hazard rates of disease progression. These findings can help nephrologists to understand the kidney disease process and better predict graft survival.

Predicting hospital-acquired infections by scoring system with simple parameters

Background

Hospital-acquired infections (HAI) are associated with increased attributable morbidity, mortality, prolonged hospitalization, and economic costs. A simple, reliable prediction model for HAI has great clinical relevance. The objective of this study is to develop a scoring system to predict HAI that was derived from Logistic Regression (LR) and validated by Artificial Neural Networks (ANN) simultaneously.

Methodology/Principal Findings

A total of 476 patients from all the 806 HAI inpatients were included for the study between 2004 and 2005. A sample of 1,376 non-HAI inpatients was randomly drawn from all the admitted patients in the same period of time as the control group. External validation of 2,500 patients was abstracted from another academic teaching center. Sixteen variables were extracted from the Electronic Health Records (EHR) and fed into ANN and LR models. With stepwise selection, the following seven variables were identified by LR models as statistically significant: Foley catheterization, central venous catheterization, arterial line, nasogastric tube, hemodialysis, stress ulcer prophylaxes and systemic glucocorticosteroids. Both ANN and LR models displayed excellent discrimination (area under the receiver operating characteristic curve [AUC]: 0.964 versus 0.969, p = 0.507) to identify infection in internal validation. During external validation, high AUC was obtained from both models (AUC: 0.850 versus 0.870, p = 0.447). The scoring system also performed extremely well in the internal (AUC: 0.965) and external (AUC: 0.871) validations.

Conclusions

We developed a scoring system to predict HAI with simple parameters validated with ANN and LR models. Armed with this scoring system, infectious disease specialists can more efficiently identify patients at high risk for HAI during hospitalization. Further, using parameters either by observation of medical devices used or data obtained from EHR also provided good prediction outcome that can be utilized in different clinical settings.

A flexible semi-Markov model for interval-censored data and goodness-of-fit testing

Multi-state approaches are becoming increasingly popular to analyse the complex evolution of patients with chronic diseases. For example, the evolution of kidney transplant recipients can be broken down into several clinical states. With this application in mind, we present a flexible semi-Markov model. The distribution functions are fitted to the durations in states and the relevance of the generalised Weibull distribution is shown. The corresponding likelihood function allows for interval censoring, i.e. the times of transitions and the sequences of states are not available during the elapsed times between two visits. The explanatory variables are introduced through the Markov chain and through the probability density functions of durations. A goodness-of-fit test is also defined to examine the stationarity of the semi-Markov model.

Hydrocortisone therapy for patients with septic shock

BACKGROUND

Hydrocortisone is widely used in patients with septic shock even though a survival benefit has been reported only in patients who remained hypotensive after fluid and vasopressor resuscitation and whose plasma cortisol levels did not rise appropriately after the administration of corticotropin.

METHODS

In this multicenter, randomized, double-blind, placebo-controlled trial, we assigned 251 patients to receive 50 mg of intravenous hydrocortisone and 248 patients to receive placebo every 6 hours for 5 days; the dose was then tapered during a 6-day period. At 28 days, the primary outcome was death among patients who did not have a response to a corticotropin test.

RESULTS

Of the 499 patients in the study, 233 (46.7%) did not have a response to corticotropin (125 in the hydrocortisone group and 108 in the placebo group). At 28 days, there was no significant difference in mortality between patients in the two study groups who did not have a response to corticotropin (39.2% in the hydrocortisone group and 36.1% in the placebo group, P=0.69) or between those who had a response to corticotropin (28.8% in the hydrocortisone group and 28.7% in the placebo group, P=1.00). At 28 days, 86 of 251 patients in the hydrocortisone group (34.3%) and 78 of 248 patients in the placebo group (31.5%) had died (P=0.51). In the hydrocortisone group, shock was reversed more quickly than in the placebo group. However, there were more episodes of superinfection, including new sepsis and septic shock.

CONCLUSIONS

Hydrocortisone did not improve survival or reversal of shock in patients with septic shock, either overall or in patients who did not have a response to corticotropin, although hydrocortisone hastened reversal of shock in patients in whom shock was reversed. (ClinicalTrials.gov number, NCT00147004.)

Attributable fractions with survival data

Attributable fraction (AF) is an important concept in clinical and epidemiological studies. The concept has mainly been discussed in relation to case-control studies, cross-sectional studies, and follow-up studies of fixed length. Here, we propose and discuss several ways of defining and estimating AFs with right-censored survival data, and thus with varying lengths of follow-up. In particular, we define the attributable hazard fraction, the AF before time t, and the AF within study. These measures have different interpretations and may give different numerical values, as illustrated in an application to real data on time to the first receiving of cash benefits for hearing impairment in children. The results underline the need for careful selection of the type of measure and interpretation when reporting AFs for survival data.

Disease progression and survival in ALS: first multi-state model approach

Although several prognostic factors have been identified in ALS, there remains some discordance concerning the prognostic significance of the age and clinical form at onset. In order to clarify these findings, we have analysed already known prognostic factors using a multi-state model. Two hundred and twenty-two sporadic ALS patients were followed. A simple unidirectional three-states model was used to summarize clinical course of ALS. States 1 and 2 reflected the progression of neurological impairment and state 3 represented the end of follow-up (tracheotomy or death). Gender, diagnostic delay, body mass index (BMI) and slow vital capacity (SVC) were also recorded. A time-inhomogeneous Markov model with piecewise constant transition intensities was used to estimate the effect of the covariates in each transition. The bulbar form at onset was only correlated with a more rapid clinical progression between state 1 and state 2. In contrast, an advanced age at diagnosis affected only survival from state 2. This methodological approach suggests that these two factors have a different prognostic significance: age at onset is related to patient's survival and the clinical form at onset predicts the progression of motoneuronal impairment in different regions.

An individual prediction of the future (disease-free) survival of patients with a history of early-stage cervical cancer, multistate model

To evaluate the possibility to give a prediction of the future (disease-free) survival, given the fact that a patient with a history of early-stage cervical cancer has been disease free for a specific period after treatment. Between January 1984 and April 2005, 615 patients with cervical cancer stages I-IIA underwent radical hysterectomy with or without adjuvant radiotherapy. The Kaplan-Meier method was used to detect statistical significance and multistate risk models to estimate the influence of covariates and to generate predicted survival curves by simulation. Simulations were done for patients with positive lymph nodes (n= 123), patients with negative lymph nodes (n= 492), and 4 hypothetical patients. The 5-year cancer-specific survival and disease-free survival of the entire group was 84% and 76%, respectively. The probability of death of the two lymph node groups and the four hypothetical patients was demonstrated in predicted cumulative probability plots. It is possible with multistate risk models to give a detailed prediction of the future (disease-free) survival, given the fact that a patient has been disease free for a specific period after treatment. This possibility is an important step forward to improve the quality of cancer care.

Predicting Hospital Mortality, Length of Stay, and Transfer to Long-Term Care for Injured Patients

BACKGROUND

Using hospital length of stay (LOS) to measure trauma care efficiency is complicated by short LOS resulting from early mortality or transfer to long-term care (LTC).

METHODS

Records from the 1999 to 2003 National Trauma Data Bank were used to create a multistate model divided into four time periods, each with constant rates of death, discharge home, and LTC transfer. Estimated hospital mortality and LOS for patient subgroups were calculated from this model, and time-varying covariate effects were estimated.

RESULTS

A total of 369,829 cases with adequate data were available. Early mortality was increased most by hypotension or coma, and also by anatomic injury severity or penetrating mechanism, but these effects diminished with time; age remained a strong predictor of mortality at any time but sex was insignificant. Rates of discharge home decreased with time, whereas rates of LTC transfer peaked at 6 to 11 days. Increased age strongly predicted transfer to LTC, whereas penetrating or burn mechanisms made it less likely. Predicted and observed outcomes were similar for multiple subgroups, and about 17% of individual variation in LOS was explained by the model.

CONCLUSIONS

Multistate models of patient status can accurately predict mortality and resource use after injury, and describe time-varying effects of other factors.

A Markov computer simulation model of the economics of neuromuscular blockade in patients with acute respiratory distress syndrome

Background

Management of acute respiratory distress syndrome (ARDS) in the intensive care unit (ICU) is clinically challenging and costly. Neuromuscular blocking agents may facilitate mechanical ventilation and improve oxygenation, but may result in prolonged recovery of neuromuscular function and acute quadriplegic myopathy syndrome (AQMS). The goal of this study was to address a hypothetical question via computer modeling: Would a reduction in intubation time of 6 hours and/or a reduction in the incidence of AQMS from 25% to 21%, provide enough benefit to justify a drug with an additional expenditure of $267 (the difference in acquisition cost between a generic and brand name neuromuscular blocker)?

Methods

The base case was a 55 year-old man in the ICU with ARDS who receives neuromuscular blockade for 3.5 days. A Markov model was designed with hypothetical patients in 1 of 6 mutually exclusive health states: ICU-intubated, ICU-extubated, hospital ward, long-term care, home, or death, over a period of 6 months. The net monetary benefit was computed.

Results

Our computer simulation modeling predicted the mean cost for ARDS patients receiving standard care for 6 months to be $62,238 (5% – 95% percentiles $42,259 – $83,766), with an overall 6-month mortality of 39%. Assuming a ceiling ratio of $35,000, even if a drug (that cost $267 more) hypothetically reduced AQMS from 25% to 21% and decreased intubation time by 6 hours, the net monetary benefit would only equal $137.

Conclusion

ARDS patients receiving a neuromuscular blocker have a high mortality, and unpredictable outcome, which results in large variability in costs per case. If a patient dies, there is no benefit to any drug that reduces ventilation time or AQMS incidence. A prospective, randomized pharmacoeconomic study of neuromuscular blockers in the ICU to asses AQMS or intubation times is impractical because of the highly variable clinical course of patients with ARDS.

Use of stochastic epidemic modeling to quantify transmission rates of colonization with methicillin-resistant Staphylococcus aureus in an intensive care unit

OBJECTIVE

To consider statistical methods for estimating transmission rates for colonization of patients with methicillin-resistant Staphylococcus aureus (MRSA) in an intensive care unit (ICU) from three different sources: background contamination, non-isolated patients, and isolated patients.

METHODS

We developed statistical methods that allowed for the analysis of interval-censored, routine surveillance data and extended the general epidemic model for the flow of patients through the ICU.

RESULTS

Within this ICU, the rate of transmission to susceptible patients from a background source of MRSA (0.0092 case per day; 95% confidence interval [CI95], 0.0062-0.0126) is approximately double the rate of transmission from a non-isolated patient (0.0052 case per day; CI95, 0.0013-0.0096) and six times the rate of transmission from an isolated patient (0.0015 case per day; CI95, 0.0001-0.0043). We used the methodology to investigate whether transmission rates vary with workload.

CONCLUSION

Our methodology has general application to infection by and transmission of pathogens in a hospital setting and is appropriate for quantifying the effect of infection control interventions.

Joint modelling of repeated transitions in follow-up data--a case study on breast cancer data

In longitudinal studies where time to a final event is the ultimate outcome often information is available about intermediate events the individuals may experience during the observation period. Even though many extensions of the Cox proportional hazards model have been proposed to model such multivariate time-to-event data these approaches are still very rarely applied to real datasets. The aim of this paper is to illustrate the application of extended Cox models for multiple time-to-event data and to show their implementation in popular statistical software packages. We demonstrate a systematic way of jointly modelling similar or repeated transitions in follow-up data by analysing an event-history dataset consisting of 270 breast cancer patients, that were followed-up for different clinical events during treatment in metastatic disease. First, we show how this methodology can also be applied to non Markovian stochastic processes by representing these processes as "conditional" Markov processes. Secondly, we compare the application of different Cox-related approaches to the breast cancer data by varying their key model components (i.e. analysis time scale, risk set and baseline hazard function). Our study showed that extended Cox models are a powerful tool for analysing complex event history datasets since the approach can address many dynamic data features such as multiple time scales, dynamic risk sets, time-varying covariates, transition by covariate interactions, autoregressive dependence or intra-subject correlation.

Ventilator-associated pulmonary infection: the germ theory of disease remains viable

Pulmonary infection complicating mechanical ventilation is a major problem in critical care. The key issues surrounding care of patients suspected of having this disease are 1) appropriate diagnostic criteria; 2) when antibiotic therapy should be started; and 3) what constitutes adequate antibiotic therapy. Current data support use of quantitative cultures obtained by either bronchoscopic or blind catheter lavage or mini-brushing. Antibiotic therapy should be guided by duration of hospitalization prior to presumed infection and local predominating nosocomial organisms and their microbial resistance patterns. The key issue with timing of therapy now centers around early termination of therapy if quantitative cultures are negative.

Reassessing the value of short-term mortality in sepsis: Comparing conventional approaches to modeling

OBJECTIVE

Clinical trials of therapies for sepsis have been mostly unsuccessful in impacting mortality. This may be partly due to the use of insensitive mortality end points. We explored whether modeling survival was more sensitive than traditional end points in detecting mortality differences in cohorts of patients with sepsis.

DESIGN

Patients were stratified into seven a priori defined paired subgroups that reflected high and low mortality risk according to known clinical risk factors. We fitted an exponential survival model to the high- and low-risk cohort of each subgroup, providing estimates of the rate of dying, long-term survival, and excess day 1 mortality. Mortality in the high- and low-risk cohorts in each subgroup was compared using model parameters, fixed-point mortality, and Kaplan-Meier survival analysis.

SETTING

Eight intensive care units within a university teaching institution.

PATIENTS

One hundred thirty patients with severe sepsis or suspected Gram-negative bacteremia.

INTERVENTIONS

None.

MEASUREMENTS AND MAIN RESULTS

Overall mortality of the cohort was 58.5% at 28 days. The survival of the entire cohort was well described by an exponential model (r2 =.99). Modeling identified differences in high- and low-risk cohorts in five of the seven paired subgroups, while conventional end-points only detected differences in 2.

CONCLUSIONS

Modeling survival was more sensitive than conventional end-points in identifying survival differences between high- and low-risk subgroups. We encourage further evaluation of modeling in the search for more sensitive mortality end points.