Development and evaluation of regression tree models for predicting in-hospital mortality of a national registry of COVID-19 patients over six pandemic surges.

BACKGROUND: Objective prognostic information is essential for good clinical decision making. In case of unknown diseases, scarcity of evidence and limited tacit knowledge prevent obtaining this information. Prediction models can be useful, but need to be not only evaluated on how well they predict, but also how stable these models are under fast changing circumstances with respect to development of the disease and the corresponding clinical response. This study aims to provide interpretable and actionable insights, particularly for clinicians. We developed and evaluated two regression tree predictive models for in-hospital mortality of COVID-19 patient at admission and 24 hours (24 h) after admission, using a national registry. We performed a retrospective analysis of observational routinely collected data. METHODS: Two regression tree models were developed for admission and 24 h after admission. The complexity of the trees was managed via cross validation to prevent overfitting. The predictive ability of the model was assessed via bootstrapping using the Area under the Receiver-Operating-Characteristic curve, Brier score and calibration curves. The tree models were assessed on the stability of their probabilities and predictive ability, on the selected variables, and compared to a full-fledged logistic regression model that uses variable selection and variable transformations using splines. Participants included COVID-19 patients from all ICUs participating in the Dutch National Intensive Care Evaluation (NICE) registry, who were admitted at the ICU between February 27, 2020, and November 23, 2021. From the NICE registry, we included concerned demographic data, minimum and maximum values of physiological data in the first 24 h of ICU admission and diagnoses (reason for admission as well as comorbidities) for model development. The main outcome measure was in-hospital mortality. We additionally analysed the Length-of-Stay (LoS) per patient subgroup per survival status. RESULTS: A total of 13,369 confirmed COVID-19 patients from 70 ICUs were included (with mortality rate of 28%). The optimism-corrected AUROC of the admission tree (with seven paths) was 0.72 (95% CI: 0.71-0.74) and of the 24 h tree (with 11 paths) was 0.74 (0.74-0.77). Both regression trees yielded good calibration and variable selection for both trees was stable. Patient subgroups comprising the tree paths had comparable survival probabilities as the full-fledged logistic regression model, survival probabilities were stable over six COVID-19 surges, and subgroups were shown to have added predictive value over the individual patient variables. CONCLUSIONS: We developed and evaluated regression trees, which operate at par with a carefully crafted logistic regression model. The trees consist of homogenous subgroups of patients that are described by simple interpretable constraints on patient characteristics thereby facilitating shared decision-making.

Prognostic models of in-hospital mortality of intensive care patients using neural representation of unstructured text: A systematic review and critical appraisal.

OBJECTIVE: To review and critically appraise published and preprint reports of prognostic models of in-hospital mortality of patients in the intensive-care unit (ICU) based on neural representations (embeddings) of clinical notes. METHODS: PubMed and arXiv were searched up to August 1, 2022. At least two reviewers independently selected the studies that developed a prognostic model of in-hospital mortality of intensive-care patients using free-text represented as embeddings and extracted data using the CHARMS checklist. Risk of bias was assessed using PROBAST. Reporting on the model was assessed with the TRIPOD guideline. To assess the machine learning components that were used in the models, we present a new descriptive framework based on different techniques to represent text and provide predictions from text. The study protocol was registered in the PROSPERO database (CRD42022354602). RESULTS: Eighteen studies out of 2,825 were included. All studies used the publicly-available MIMIC dataset. Context-independent word embeddings are widely used. Model discrimination was provided by all studies (AUROC 0.75-0.96), but measures of calibration were scarce. Seven studies used both structural clinical variables and notes. Model discrimination improved when adding clinical notes to variables. None of the models was externally validated and often a simple train/test split was used for internal validation. Our critical appraisal demonstrated a high risk of bias in all studies and concerns regarding their applicability in clinical practice. CONCLUSION: All studies used a neural architecture for prediction and were based on one publicly available dataset. Clinical notes were reported to improve predictive performance when used in addition to only clinical variables. Most studies had methodological, reporting, and applicability issues. We recommend reporting both model discrimination and calibration, using additional data sources, and using more robust evaluation strategies, including prospective and external validation. Finally, sharing data and code is encouraged to improve study reproducibility.

Automated identification of patient subgroups: A case-study on mortality of COVID-19 patients admitted to the ICU.

BACKGROUND: - Subgroup discovery (SGD) is the automated splitting of the data into complex subgroups. Various SGD methods have been applied to the medical domain, but none have been extensively evaluated. We assess the numerical and clinical quality of SGD methods. METHOD: - We applied the improved Subgroup Set Discovery (SSD++), Patient Rule Induction Method (PRIM) and APRIORI - Subgroup Discovery (APRIORI-SD) algorithms to obtain patient subgroups on observational data of 14,548 COVID-19 patients admitted to 73 Dutch intensive care units. Hospital mortality was the clinical outcome. Numerical significance of the subgroups was assessed with information-theoretic measures. Clinical significance of the subgroups was assessed by comparing variable importance on population and subgroup levels and by expert evaluation. RESULTS: - The tested algorithms varied widely in the total number of discovered subgroups (5-62), the number of selected variables, and the predictive value of the subgroups. Qualitative assessment showed that the found subgroups make clinical sense. SSD++ found most subgroups (n = 62), which added predictive value and generally showed high potential for clinical use. APRIORI-SD and PRIM found fewer subgroups (n = 5 and 6), which did not add predictive value and were clinically less relevant. CONCLUSION: - Automated SGD methods find clinical subgroups that are relevant when assessed quantitatively (yield added predictive value) and qualitatively (intensivists consider the subgroups significant). Different methods yield different subgroups with varying degrees of predictive performance and clinical quality. External validation is needed to generalize the results to other populations and future research should explore which algorithm performs best in other settings.

Identifying prognostic factors for clinical outcomes and costs in four high-volume surgical treatments using routinely collected hospital data.

Identifying prognostic factors (PFs) is often costly and labor-intensive. Routinely collected hospital data provide opportunities to identify clinically relevant PFs and construct accurate prognostic models without additional data-collection costs. This multicenter (66 hospitals) study reports on associations various patient-level variables have with outcomes and costs. Outcomes were in-hospital mortality, intensive care unit (ICU) admission, length of stay, 30-day readmission, 30-day reintervention and in-hospital costs. Candidate PFs were age, sex, Elixhauser Comorbidity Score, prior hospitalizations, prior days spent in hospital, and socio-economic status. Included patients dealt with either colorectal carcinoma (CRC, n = 10,254), urinary bladder carcinoma (UBC, n = 17,385), acute percutaneous coronary intervention (aPCI, n = 25,818), or total knee arthroplasty (TKA, n = 39,214). Prior hospitalization significantly increased readmission risk in all treatments (OR between 2.15 and 25.50), whereas prior days spent in hospital decreased this risk (OR between 0.55 and 0.95). In CRC patients, women had lower risk of in-hospital mortality (OR 0.64), ICU admittance (OR 0.68) and 30-day reintervention (OR 0.70). Prior hospitalization was the strongest PF for higher costs across all treatments (31-64% costs increase/hospitalization). Prognostic model performance (c-statistic) ranged 0.67-0.92, with Brier scores below 0.08. R-squared ranged from 0.06-0.19 for LoS and 0.19-0.38 for costs. Identified PFs should be considered as building blocks for treatment-specific prognostic models and information for monitoring patients after surgery. Researchers and clinicians might benefit from gaining a better insight into the drivers behind (costs) prognosis.

Can we reliably automate clinical prognostic modelling? A retrospective cohort study for ICU triage prediction of in-hospital mortality of COVID-19 patients in the Netherlands.

BACKGROUND: Building Machine Learning (ML) models in healthcare may suffer from time-consuming and potentially biased pre-selection of predictors by hand that can result in limited or trivial selection of suitable models. We aimed to assess the predictive performance of automating the process of building ML models (AutoML) in-hospital mortality prediction modelling of triage COVID-19 patients at ICU admission versus expert-based predictor pre-selection followed by logistic regression. METHODS: We conducted an observational study of all COVID-19 patients admitted to Dutch ICUs between February and July 2020. We included 2,690 COVID-19 patients from 70 ICUs participating in the Dutch National Intensive Care Evaluation (NICE) registry. The main outcome measure was in-hospital mortality. We asessed model performance (at admission and after 24h, respectively) of AutoML compared to the more traditional approach of predictor pre-selection and logistic regression. FINDINGS: Predictive performance of the autoML models with variables available at admission shows fair discrimination (average AUROC = 0·75-0·76 (sdev = 0·03), PPV = 0·70-0·76 (sdev = 0·1) at cut-off = 0·3 (the observed mortality rate), and good calibration. This performance is on par with a logistic regression model with selection of patient variables by three experts (average AUROC = 0·78 (sdev = 0·03) and PPV = 0·79 (sdev = 0·2)). Extending the models with variables that are available at 24h after admission resulted in models with higher predictive performance (average AUROC = 0·77-0·79 (sdev = 0·03) and PPV = 0·79-0·80 (sdev = 0·10-0·17)). CONCLUSIONS: AutoML delivers prediction models with fair discriminatory performance, and good calibration and accuracy, which is as good as regression models with expert-based predictor pre-selection. In the context of the restricted availability of data in an ICU quality registry, extending the models with variables that are available at 24h after admission showed small (but significantly) performance increase.

Interaction between geriatric syndromes in predicting three months mortality risk.

OBJECTIVES: Capturing frailty using a quick tool has proven to be challenging. We hypothesise that this is due to the complex interactions between frailty domains. We aimed to identify these interactions and assess whether adding interactions between domains improves mortality predictability. METHODS: In this retrospective cohort study, we selected all patients aged 70 or older who were admitted to one Dutch hospital between April 2015 and April 2016. Patient characteristics, frailty screening (using VMS (Safety Management System), a screening tool used in Dutch hospital care), length of stay, and mortality within three months were retrospectively collected from electronic medical records. To identify predictive interactions between the frailty domains, we constructed a classification tree with mortality as the outcome using five variables: the four VMS-domains (delirium risk, fall risk, malnutrition, physical impairment) and their sum. To determine if any domain interactions were predictive for three-month mortality, we performed a multivariable logistic regression analysis. RESULTS: We included 4,478 patients. (median age: 79 years; maximum age: 101 years; 44.8% male) The highest risk for three-month mortality included patients that were physically impaired and malnourished (23% (95%-CI 19.0-27.4%)). Subgroups had comparable three-month mortality risks based on different domains: malnutrition without physical impairment (15.2% (96%-CI 12.4-18.6%)) and physical impairment and delirium risk without malnutrition (16.3% (95%-CI 13.7-19.2%)). DISCUSSION: We showed that taking interactions between domains into account improves the predictability of three-month mortality risk. Therefore, when screening for frailty, simply adding up domains with a cut-off score results in loss of valuable information.

Explaining heterogeneity of individual treatment causal effects by subgroup discovery: An observational case study in antibiotics treatment of acute rhino-sinusitis.

OBJECTIVES: Individuals may respond differently to the same treatment, and there is a need to understand such heterogeneity of causal individual treatment effects. We propose and evaluate a modelling approach to better understand this heterogeneity from observational studies by identifying patient subgroups with a markedly deviating response to treatment. We illustrate this approach in a primary care case-study of antibiotic (AB) prescription on recovery from acute rhino-sinusitis (ARS). METHODS: Our approach consists of four stages and is applied to a large dataset in primary care dataset of 24,392 patients suspected of suffering from ARS. We first identify pre-treatment variables that either confound the relationship between treatment and outcome or are risk factors of the outcome. Second, based on the pre-treatment variables we create Synthetic Random Forest (SRF) models to compute the potential outcomes and subsequently the causal individual treatment effect (ITE) estimates. Third, we perform subgroup discovery using the ITE estimates as outcomes to identify positive and negative responders. Fourth, we evaluate the predictive performance of the identified subgroups for predicting the outcome in two ways: the likelihood ratio test, and whether the subgroups are selected via the Akaike Information Criterion (AIC) using backward stepwise variable selection. We validate the whole modelling strategy by means of 10-fold-cross-validation. RESULTS: Based on 20 pre-treatment variables, four subgroups (three for positive responders and one for negative responders) were identified. The log likelihood ratio tests showed that the subgroups were significant. Variable selection using the AIC kept two of the four subgroups, one for positive responders and one for negative responders. As for the validation of the whole modelling strategy, all reported measures (the number of pre-treatment variables associated with the outcome, number of subgroups, number of subgroups surviving variable selection and coverage) showed little variation. CONCLUSIONS: With the proposed approach, we identified subgroups of positive and negative responders to treatment that markedly deviate from the mean response. The subgroups showed additive predictive value of the outcome. The modelling approach strategy was shown to be robust on this dataset. Our approach was thus able to discover understandable subgroups from observational data that have predictive value and which may be considered by the clinical users to get insight into who responds positively or negatively to a proposed treatment.