A scalable approach for critical care data extraction and analysis in an academic medical center.

BACKGROUND: Electronic health records are a valuable asset for research, but their use is challenging due to inconsistencies of records, heterogeneous formats and the distribution over multiple, non-integrated information systems. Hence, specialized health data engineering and data science expertise are required to enable research. To facilitate secondary use of clinical routine data collected in our intensive care wards, we developed a scalable approach, consisting of cohort generation, variable filtering and data extraction steps. OBJECTIVE: With this report we share our workflow of data request, cohort identification and data extraction. We present an algorithm for automatic data extraction from our critical care information system (CCIS) that can be adapted to other object-oriented data bases. METHODS: We introduced a data request process with functionalities for automated identification of patient cohorts and a specialized hierarchical data structure that supports filtering relevant variables from the CCIS and further systems for the specified cohorts. The data extraction algorithm takes patient pseudonyms and variable lists as inputs. Algorithms are implemented in Python, leveraging the PySpark framework running on our data lake infrastructure. RESULTS: Our data request process is in operational use since June 2022. Since then we have served 121 projects with 148 service requests in total. We discuss the hierarchical structure and the frequently used data items of our CCIS in detail and present an application example, including cohort selection, data extraction and data transformation into an analyses-ready format. CONCLUSIONS: Using clinical routine data for secondary research is challenging and requires an interdisciplinary team. We developed a scalable approach that automates steps for cohort identification, data extraction and common data pre-processing steps. Additionally, we facilitate data harmonization, integration and consult on typical data analysis scenarios, machine learning algorithms and visualizations in dashboards.

Predicting postoperative delirium assessed by the Nursing Screening Delirium Scale in the recovery room for non-cardiac surgeries without craniotomy: A retrospective study using a machine learning approach.

Postoperative delirium (POD) contributes to severe outcomes such as death or development of dementia. Thus, it is desirable to identify vulnerable patients in advance during the perioperative phase. Previous studies mainly investigated risk factors for delirium during hospitalization and further used a linear logistic regression (LR) approach with time-invariant data. Studies have not investigated patients' fluctuating conditions to support POD precautions. In this single-center study, we aimed to predict POD in a recovery room setting with a non-linear machine learning (ML) technique using pre-, intra-, and postoperative data. The target variable POD was defined with the Nursing Screening Delirium Scale (Nu-DESC) ≥ 1. Feature selection was conducted based on robust univariate test statistics and L1 regularization. Non-linear multi-layer perceptron (MLP) as well as tree-based models were trained and evaluated-with the receiver operating characteristics curve (AUROC), the area under precision recall curve (AUPRC), and additional metrics-against LR and published models on bootstrapped testing data. The prevalence of POD was 8.2% in a sample of 73,181 surgeries performed between 2017 and 2020. Significant univariate impact factors were the preoperative ASA status (American Society of Anesthesiologists physical status classification system), the intraoperative amount of given remifentanil, and the postoperative Aldrete score. The best model used pre-, intra-, and postoperative data. The non-linear boosted trees model achieved a mean AUROC of 0.854 and a mean AUPRC of 0.418 outperforming linear LR, well as best applied and retrained baseline models. Overall, non-linear machine learning models using data from multiple perioperative time phases were superior to traditional ones in predicting POD in the recovery room. Class imbalance was seen as a main impediment for model application in clinical practice.

A Recurrent Neural Network Model for Predicting Activated Partial Thromboplastin Time After Treatment With Heparin: Retrospective Study.

BACKGROUND: Anticoagulation therapy with heparin is a frequent treatment in intensive care units and is monitored by activated partial thromboplastin clotting time (aPTT). It has been demonstrated that reaching an established anticoagulation target within 24 hours is associated with favorable outcomes. However, patients respond to heparin differently and reaching the anticoagulation target can be challenging. Machine learning algorithms may potentially support clinicians with improved dosing recommendations. OBJECTIVE: This study evaluates a range of machine learning algorithms on their capability of predicting the patients' response to heparin treatment. In this analysis, we apply, for the first time, a model that considers time series. METHODS: We extracted patient demographics, laboratory values, dialysis and extracorporeal membrane oxygenation treatments, and scores from the hospital information system. We predicted the numerical values of aPTT laboratory values 24 hours after continuous heparin infusion and evaluated 7 different machine learning models. The best-performing model was compared to recently published models on a classification task. We considered all data before and within the first 12 hours of continuous heparin infusion as features and predicted the aPTT value after 24 hours. RESULTS: The distribution of aPTT in our cohort of 5926 hospital admissions was highly skewed. Most patients showed aPTT values below 75 s, while some outliers showed much higher aPTT values. A recurrent neural network that consumes a time series of features showed the highest performance on the test set. CONCLUSIONS: A recurrent neural network that uses time series of features instead of only static and aggregated features showed the highest performance in predicting aPTT after heparin treatment.

MIMIC-IV as a Clinical Data Schema.

Routinely collected electronic health records (EHR) in clinical information systems (CIS) are often heterogeneous, have inconsistent data formats and lack of documentation. We use the well-known open-source database schema of MIMIC-IV to address this issue aiming to support collaborative secondary analysis. Over 154 million data records from a German ICU have already been mapped and inserted into the schema successfully. However, discrepancies between the German and US health systems as well as specifics in our clinical source data hinder the direct translation to MIMIC. Evaluating and improving mapping completeness is part of the ongoing research.

The influence of patient characteristics on the alarm rate in intensive care units: a retrospective cohort study.

Intensive care units (ICU) are often overflooded with alarms from monitoring devices which constitutes a hazard to both staff and patients. To date, the suggested solutions to excessive monitoring alarms have remained on a research level. We aimed to identify patient characteristics that affect the ICU alarm rate with the goal of proposing a straightforward solution that can easily be implemented in ICUs. Alarm logs from eight adult ICUs of a tertiary care university-hospital in Berlin, Germany were retrospectively collected between September 2019 and March 2021. Adult patients admitted to the ICU with at least 24 h of continuous alarm logs were included in the study. The sum of alarms per patient per day was calculated. The median was 119. A total of 26,890 observations from 3205 patients were included. 23 variables were extracted from patients' electronic health records (EHR) and a multivariable logistic regression was performed to evaluate the association of patient characteristics and alarm rates. Invasive blood pressure monitoring (adjusted odds ratio (aOR) 4.68, 95%CI 4.15-5.29, p < 0.001), invasive mechanical ventilation (aOR 1.24, 95%CI 1.16-1.32, p < 0.001), heart failure (aOR 1.26, 95%CI 1.19-1.35, p < 0.001), chronic renal failure (aOR 1.18, 95%CI 1.10-1.27, p < 0.001), hypertension (aOR 1.19, 95%CI 1.13-1.26, p < 0.001), high RASS (aOR 1.22, 95%CI 1.18-1.25, p < 0.001) and scheduled surgical admission (aOR 1.22, 95%CI 1.13-1.32, p < 0.001) were significantly associated with a high alarm rate. Our study suggests that patient-specific alarm management should be integrated in the clinical routine of ICUs. To reduce the overall alarm load, particular attention regarding alarm management should be paid to patients with invasive blood pressure monitoring, invasive mechanical ventilation, heart failure, chronic renal failure, hypertension, high RASS or scheduled surgical admission since they are more likely to have a high contribution to noise pollution, alarm fatigue and hence compromised patient safety in ICUs.

Predicting lethal courses in critically ill COVID-19 patients using a machine learning model trained on patients with non-COVID-19 viral pneumonia.

In a pandemic with a novel disease, disease-specific prognosis models are available only with a delay. To bridge the critical early phase, models built for similar diseases might be applied. To test the accuracy of such a knowledge transfer, we investigated how precise lethal courses in critically ill COVID-19 patients can be predicted by a model trained on critically ill non-COVID-19 viral pneumonia patients. We trained gradient boosted decision tree models on 718 (245 deceased) non-COVID-19 viral pneumonia patients to predict individual ICU mortality and applied it to 1054 (369 deceased) COVID-19 patients. Our model showed a significantly better predictive performance (AUROC 0.86 [95% CI 0.86-0.87]) than the clinical scores APACHE2 (0.63 [95% CI 0.61-0.65]), SAPS2 (0.72 [95% CI 0.71-0.74]) and SOFA (0.76 [95% CI 0.75-0.77]), the COVID-19-specific mortality prediction models of Zhou (0.76 [95% CI 0.73-0.78]) and Wang (laboratory: 0.62 [95% CI 0.59-0.65]; clinical: 0.56 [95% CI 0.55-0.58]) and the 4C COVID-19 Mortality score (0.71 [95% CI 0.70-0.72]). We conclude that lethal courses in critically ill COVID-19 patients can be predicted by a machine learning model trained on non-COVID-19 patients. Our results suggest that in a pandemic with a novel disease, prognosis models built for similar diseases can be applied, even when the diseases differ in time courses and in rates of critical and lethal courses.