Survival prediction in intensive-care units based on aggregation of long-term disease history and acute physiology: a retrospective study of the Danish National Patient Registry and electronic patient records

Evaluation of score-based tertiary triage policies during the COVID-19 pandemic: simulation study with real-world intensive care data

OBJECTIVE

The explicit prohibition of discontinuing intensive care unit (ICU) treatment that has already begun by the newly established German Triage Act in favor of new patients with better prognoses (tertiary triage) under crisis conditions may prevent saving as many patients as possible and therefore may violate the international well-accepted premise of undertaking the "best for the most" patients. During the COVID-19 pandemic, authorities set up lockdown measures and infection-prevention strategies to avoid an overburdened health-care system. In cases of situational overload of ICU resources, when transporting options are exhausted, the question of a tertiary triage of patients arises.

METHODS

We provide data-driven analyses of score- and non-score-based tertiary triage policies using simulation and real-world electronic health record data in a COVID-19 setting. Ten different triage policies, for example, based on the Simplified Acute Physiology Score (SAPS II), are compared based on the resulting mortality in the ICU and inferential statistics.

RESULTS

Our study shows that score-based tertiary triage policies outperform non-score-based tertiary triage policies including compliance with the German Triage Act. Based on our simulation model, a SAPS II score-based tertiary triage policy reduces mortality in the ICU by up to 18 percentage points. The longer the queue of critical care patients waiting for ICU treatment and the larger the maximum number of patients subject to tertiary triage, the greater the effect on the reduction of mortality in the ICU.

CONCLUSION

A SAPS II score-based tertiary triage policy was superior in our simulation model. Random allocation or "first come, first served" policies yield the lowest survival rates, as will adherence to the new German Triage Act. An interdisciplinary discussion including an ethical and legal perspective is important for the social interpretation of our data-driven results.

Disease Trajectories from Healthcare Data: Methodologies, Key Results, and Future Perspectives

Disease trajectories, defined as sequential, directional disease associations, have become an intense research field driven by the availability of electronic population-wide healthcare data and sufficient computational power. Here, we provide an overview of disease trajectory studies with a focus on European work, including ontologies used as well as computational methodologies for the construction of disease trajectories. We also discuss different applications of disease trajectories from descriptive risk identification to disease progression, patient stratification, and personalized predictions using machine learning. We describe challenges and opportunities in the area that eventually will benefit from initiatives such as the European Health Data Space, which, with time, will make it possible to analyze data from cohorts comprising hundreds of millions of patients.

The anastomotic leak triad: preoperative patient characteristics, intraoperative risk factors, and postoperative outcomes

BACKGROUND

The aim of this study was to determine perioperative risk factors associated with anastomotic leak (AL) after minimally invasive esophagectomy (MIE) and its association with cancer recurrence and overall survival.

METHODS

This retrospective observational study of electronic health record data included patients who underwent MIE for esophageal cancer between September 2013 and July 2023 at a tertiary center. The primary outcome was AL after esophagectomy, whereas the secondary outcomes included time to cancer recurrence and overall survival. Perioperative patient factors were evaluated to determine their associations with the primary and the secondary outcomes. Propensity score-matched logistic regression assessed the associations between perioperative factors and AL. Kaplan-Meier survival curves compared cancer recurrence and overall survival by AL.

RESULTS

A total of 251 consecutive patients with esophageal cancer were included in the analysis; 15 (6%) developed AL. Anemia, hospital complications, hospital length of stay, and 30-day readmissions significantly differed from those with and without AL (P = .037, <.001, <.001, and.016, respectively). Moreover, 30- and 90-day mortality were not statistically affected by the presence of AL (P = .417 and 0.456, respectively). Logistic regression modeling showed drug history and anemia were significantly associated with AL (P = .022 and.011, respectively). The presence of AL did not significantly impact cancer recurrence or overall survival (P = .439 and.301, respectively).

CONCLUSION

The etiology of AL is multifactorial. Moreover, AL is significantly associated with drug history, preoperative anemia, hospital length of stay, and 30-day readmissions, but it was not significantly associated with 30- or 90-day mortality, cancer recurrence, or overall survival. Patients should be optimized before undergoing MIE with special consideration for correcting anemia. Ongoing research is needed to identify more modifiable risk factors to minimize AL development and its associated morbidity.

Analysis and prediction of long-term survival using a clinically applicable risk score based on the Electronic Health Record

BACKGROUND

The long-term survival of a population assigned to a hospital can be essential to anticipate, manage, and provide appropriate hospital healthcare resources or lead preventive actions for high-risk mortality individuals. In this study, we discriminate which electronic health record variables are most relevant to predict the long-term survival of a population, and apply the results to identify high-risk mortality groups.

MATERIALS AND METHODS

A prospective cohort study was conducted on a population of 113,403 individuals alive on July 1st, 2018 from the General Hospital of Castellón (Spain). Considering electronic health record patients' variables and survival days from the start date of the study, a Kaplan-Meier analysis and a multivariate Cox regression model were performed, and a risk score based on Cox coefficients was applied to predict survival over 3 years.

RESULTS

All significant covariates from the Cox model (91.5% c-index) were associated with increased mortality risk. Using the proposed risk score, Kaplan-Meier curves show that survival probability in the 3rd year is 99.23% (95% confidence interval (CI) 99.18-99.29) for the low-risk, 91.21% (95% CI 90.67-91.76) for medium-risk, 76.52% (95% CI 75.59-77.46) for the high-risk, and 48.61 % (95% CI 46.85-50.36) for the very high-risk groups.

DISCUSSION

The Cox model obtained is highly predictive, and it has been found that some electronic health record variables little studied to date, such as Clinical Risk Groups, have a strong impact on survival. Regarding clinical application, the proposed risk score is particularly useful for identifying high-risk subpopulations within a large population.

Predictive approach for liberation from acute dialysis in ICU patients using interpretable machine learning

Renal recovery following dialysis-requiring acute kidney injury (AKI-D) is a vital clinical outcome in critical care, yet it remains an understudied area. This retrospective cohort study, conducted in a medical center in Taiwan from 2015 to 2020, enrolled patients with AKI-D during intensive care unit stays. We aimed to develop and temporally test models for predicting dialysis liberation before hospital discharge using machine learning algorithms and explore early predictors. The dataset comprised 90 routinely collected variables within the first three days of dialysis initiation. Out of 1,381 patients who received acute dialysis, 27.3% experienced renal recovery. The cohort was divided into the training group (N = 1135) and temporal testing group (N = 251). The models demonstrated good performance, with an area under the receiver operating characteristic curve of 0.85 (95% CI, 0.81-0.88) and an area under the precision-recall curve of 0.69 (95% CI, 0.62-0.76) for the XGBoost model. Key predictors included urine volume, Charlson comorbidity index, vital sign derivatives (trend of respiratory rate and SpO2), and lactate levels. We successfully developed early prediction models for renal recovery by integrating early changes in vital signs and inputs/outputs, which have the potential to aid clinical decision-making in the ICU.

Multi-cancer risk stratification based on national health data: a retrospective modelling and validation study

BACKGROUND

Health care is experiencing a drive towards digitisation, and many countries are implementing national health data resources. Although a range of cancer risk models exists, the utility on a population level for risk stratification across cancer types has not been fully explored. We aimed to close this gap by evaluating pan-cancer risk models built on electronic health records across the Danish population with validation in the UK Biobank.

METHODS

In this retrospective modelling and validation study, data for model development and internal validation were derived from the following Danish health registries: the Central Person Registry, the Danish National Patient Registry, the death registry, the cancer registry, and full-text medical records from secondary care records in the capital region. The development data included adults aged 16-86 years without previous malignant cancers in the time period from Jan 1, 1995, to Dec 31, 2014. The internal validation period was from Jan 1, 2015, to April 10, 2018, and the data included all adults without a previous indication of cancer aged 16-75 years on Dec 31, 2014. The external validation cohort from the UK Biobank included all adults without a previous indication of cancer aged 50-75 years. We used time-dependent Bayesian Cox hazard models built on the combined medical history of Danish individuals. A set of 1392 covariates from available clinical disease trajectories, text-mined basic health factors, and family histories were used to train predictive models of 20 major cancer types. The models were validated on cancer incidence between 2015 and 2018 across Denmark and on individuals in the UK Biobank. The primary outcomes were discrimination and calibration performance.

FINDINGS

From the Danish registries, we included 6 732 553 individuals covering 60 million hospital visits, 90 million diagnoses, and a total of 193 million life-years between Jan 1, 1978, and April 10, 2018. Danish registry data covering the period from Jan 1, 2015, to April 10, 2018, were used to internally validate risk models, containing a total of 4 248 491 individuals who remained at risk of a primary malignant cancer diagnosis and 67 401 cancer cases recorded. For the external validation, we evaluated the same time period in the UK Biobank covering 377 004 individuals with 11 486 cancer cases. The predictive performance of the models on Danish data showed good discrimination (concordance index 0·81 [SD 0·08], ranging from 0·66 [95% CI 0·65-0·67] for cervix uteri cancer to 0·91 [0·90-0·92] for liver cancer). Performance was similar on the UK Biobank in a direct transfer when controlling for shifts in the age distribution (concordance index 0·66 [SD 0·08], ranging from 0·55 [95% CI 0·44-0·66] for cervix uteri cancer to 0·78 [0·77-0·79] for lung cancer). Cancer risks were associated, in addition to heritable components, with a broad range of preceding diagnoses and health factors. The best overall performance was seen for cancers of the digestive system (oesophageal, stomach, colorectal, liver, and pancreatic) but also thyroid, kidney, and uterine cancers.

INTERPRETATION

Data available in national electronic health databases can be used to approximate cancer risk factors and enable risk predictions in most cancer types. Model predictions generalise between the Danish and UK health-care systems. With the emergence of multi-cancer early detection tests, electronic health record-based risk models could supplement screening efforts.

FUNDING

Novo Nordisk Foundation and the Danish Innovation Foundation.

Supervised machine learning model to predict mortality in patients undergoing venovenous extracorporeal membrane oxygenation from a nationwide multicentre registry

BACKGROUND

Existing models have performed poorly when predicting mortality for patients undergoing venovenous extracorporeal membrane oxygenation (VV-ECMO). This study aimed to develop and validate a machine learning (ML)-based prediction model to predict 90-day mortality in patients undergoing VV-ECMO.

METHODS

This study included 368 patients with acute respiratory failure undergoing VV-ECMO from 16 tertiary hospitals across South Korea between 2012 and 2015. The primary outcome was the 90-day mortality after ECMO initiation. The inputs included all available features (n=51) and those from the electronic health record (EHR) systems without preprocessing (n=40). The discriminatory strengths of ML models were evaluated in both internal and external validation sets. The models were compared with conventional models, such as respiratory ECMO survival prediction (RESP) and predicting death for severe acute respiratory distress syndrome on VV-ECMO (PRESERVE).

RESULTS

Extreme gradient boosting (XGB) (areas under the receiver operating characteristic curve, AUROC 0.82, 95% CI (0.73 to 0.89)) and light gradient boosting (AUROC 0.81 (95% CI 0.71 to 0.88)) models achieved the highest performance using EHR's and all other available features. The developed models had higher AUROCs (95% CI 0.76 to 0.82) than those of RESP (AUROC 0.66 (95% CI 0.56 to 0.76)) and PRESERVE (AUROC 0.71 (95% CI 0.61 to 0.81)). Additionally, we achieved an AUROC (0.75) for 90-day mortality in external validation in the case of the XGB model, which was higher than that of RESP (0.70) and PRESERVE (0.67) in the same validation dataset.

CONCLUSIONS

ML prediction models outperformed previous mortality risk models. This model may be used to identify patients who are unlikely to benefit from VV-ECMO therapy during patient selection.

Application of machine learning in predicting survival outcomes involving real-world data: a scoping review

BACKGROUND

Despite the interest in machine learning (ML) algorithms for analyzing real-world data (RWD) in healthcare, the use of ML in predicting time-to-event data, a common scenario in clinical practice, is less explored. ML models are capable of algorithmically learning from large, complex datasets and can offer advantages in predicting time-to-event data. We reviewed the recent applications of ML for survival analysis using RWD in healthcare.

METHODS

PUBMED and EMBASE were searched from database inception through March 2023 to identify peer-reviewed English-language studies of ML models for predicting time-to-event outcomes using the RWD. Two reviewers extracted information on the data source, patient population, survival outcome, ML algorithms, and the Area Under the Curve (AUC).

RESULTS

Of 257 citations, 28 publications were included. Random survival forests (N = 16, 57%) and neural networks (N = 11, 39%) were the most popular ML algorithms. There was variability across AUC for these ML models (median 0.789, range 0.6-0.950). ML algorithms were predominately considered for predicting overall survival in oncology (N = 12, 43%). ML survival models were often used to predict disease prognosis or clinical events (N = 27, 96%) in the oncology, while less were used for treatment outcomes (N = 1, 4%).

CONCLUSIONS

The ML algorithms, random survival forests and neural networks, are mainly used for RWD to predict survival outcomes such as disease prognosis or clinical events in the oncology. This review shows that more opportunities remain to apply these ML algorithms to inform treatment decision-making in clinical practice. More methodological work is also needed to ensure the utility and applicability of ML models in survival outcomes.

A deep learning algorithm to predict risk of pancreatic cancer from disease trajectories

Pancreatic cancer is an aggressive disease that typically presents late with poor outcomes, indicating a pronounced need for early detection. In this study, we applied artificial intelligence methods to clinical data from 6 million patients (24,000 pancreatic cancer cases) in Denmark (Danish National Patient Registry (DNPR)) and from 3 million patients (3,900 cases) in the United States (US Veterans Affairs (US-VA)). We trained machine learning models on the sequence of disease codes in clinical histories and tested prediction of cancer occurrence within incremental time windows (CancerRiskNet). For cancer occurrence within 36 months, the performance of the best DNPR model has area under the receiver operating characteristic (AUROC) curve = 0.88 and decreases to AUROC (3m) = 0.83 when disease events within 3 months before cancer diagnosis are excluded from training, with an estimated relative risk of 59 for 1,000 highest-risk patients older than age 50 years. Cross-application of the Danish model to US-VA data had lower performance (AUROC = 0.71), and retraining was needed to improve performance (AUROC = 0.78, AUROC (3m) = 0.76). These results improve the ability to design realistic surveillance programs for patients at elevated risk, potentially benefiting lifespan and quality of life by early detection of this aggressive cancer.

ICD2Vec: Mathematical representation of diseases

BACKGROUND

The International Classification of Diseases (ICD) codes represent the global standard for reporting disease conditions. The current ICD codes connote direct human-defined relationships among diseases in a hierarchical tree structure. Representing the ICD codes as mathematical vectors helps to capture nonlinear relationships in medical ontologies across diseases.

METHODS

We propose a universally applicable framework called "ICD2Vec" designed to provide mathematical representations of diseases by encoding corresponding information. First, we present the arithmetical and semantic relationships between diseases by mapping composite vectors for symptoms or diseases to the most similar ICD codes. Second, we investigated the validity of ICD2Vec by comparing the biological relationships and cosine similarities among the vectorized ICD codes. Third, we propose a new risk score called IRIS, derived from ICD2Vec, and demonstrate its clinical utility with large cohorts from the UK and South Korea.

RESULTS

Semantic compositionality was qualitatively confirmed between descriptions of symptoms and ICD2Vec. For example, the most diseases most similar to COVID-19 were found to be the common cold (ICD-10: J00), unspecified viral hemorrhagic fever (ICD-10: A99), and smallpox (ICD-10: B03). We show the significant associations between the cosine similarities derived from ICD2Vec and the biological relationships using disease-to-disease pairs. Furthermore, we observed significant adjusted hazard ratios (HR) and area under the receiver operating characteristics (AUROC) between IRIS and risks for eight diseases. For instance, the higher IRIS for coronary artery disease (CAD) can be the higher probability for the incidence of CAD (HR: 2.15 [95% CI 2.02-2.28] and AUROC: 0.587 [95% CI 0.583-0.591]). We identified individuals at substantially increased risk of CAD using IRIS and 10-year atherosclerotic cardiovascular disease risk (adjusted HR, 4.26, 95% CI, 3.59-5.05).

CONCLUSIONS

ICD2Vec, a proposed universal framework for converting qualitatively measured ICD codes into quantitative vectors containing semantic relationships between diseases, exhibited a significant correlation with actual biological significance. In addition, the IRIS was a significant predictor of major diseases in a prospective study using two large-scale Biobank EHR datasets. Based on this clinical validity and utility evidence, we suggest that publicly available ICD2Vec can be used in diverse research and clinical practices and has important clinical implications.

Tendency of dynamic vasoactive and inotropic medications data as a robust predictor of mortality in patients with septic shock: An analysis of the MIMIC-IV database

Background

Septic shock patients fundamentally require delicate vasoactive and inotropic agent administration, which could be quantitatively and objectively evaluated by the vasoactive–inotropic score (VIS); however, whether the dynamic trends of high-time-resolution VIS alter the clinical outcomes remains unclear. Thus, this study proposes the term VIS Reduction Rate (VRR) to generalise the tendency of dynamic VIS, to explore the association of VRR and mortality for patients with septic shock.

Methods

We applied dynamic and static VIS data to predict ICU mortality by two models: the long short-term memory (LSTM) deep learning model, and the extreme gradient boosting (XGBoost), respectively. The specific target cohort was extracted from the Medical Information Mart for Intensive Care IV (MIMIC-IV) database by the sophisticated structured query language (SQL). Enrolled patients were divided into four groups by VRR value: ≥50%, 0 ~ 50%, −50% ~ 0, and &lt; −50%. Statistical approaches included pairwise propensity score matching (PSM), Cox proportional hazards regression, and two doubly robust estimation models to ensure the robustness of the results. The primary and secondary outcomes were ICU mortality and in-hospital mortality, respectively.

Results

VRR simplifies the dosing trends of vasoactive and inotropic agents represented by dynamic VIS data while requiring fewer data. In total, 8,887 septic shock patients were included. Compared with the VRR ≥50% group, the 0 ~ 50%, −50% ~ 0, and &lt; −50% groups had significantly higher ICU mortality [hazard ratio (HR) 1.32, 95% confidence interval (CI) 1.17–1.50, p &lt; 0.001; HR 1.79, 95% CI 1.44–2.22, p &lt; 0.001; HR 2.07, 95% CI 1.61–2.66, p &lt; 0.001, respectively] and in-hospital mortality [HR 1.43, 95% CI 1.28–1.60, p &lt; 0.001; HR 1.75, 95% CI 1.45–2.11, p &lt; 0.001; HR 2.00, 95% CI 1.61–2.49, p &lt; 0.001, respectively]. Similar findings were observed in two doubly robust estimation models.

Conclusion

The trends of dynamic VIS in ICU might help intensivists to stratify the prognosis of adult patients with septic shock. A lower decline of VIS was remarkably associated with higher ICU and in-hospital mortality among septic shock patients receiving vasoactive–inotropic therapy for more than 24 h.

Elderly in the ICU

How to cite this article: Jagathkar G. Elderly in the ICU. Indian J Crit Care Med 2023;27(3):157-158.

Deep Learning for Cross-Diagnostic Prediction of Mental Disorder Diagnosis and Prognosis Using Danish Nationwide Register and Genetic Data

Importance

Diagnoses and treatment of mental disorders are hampered by the current lack of objective markers needed to provide a more precise diagnosis and treatment strategy.

Objective

To develop deep learning models to predict mental disorder diagnosis and severity spanning multiple diagnoses using nationwide register data, family and patient-specific diagnostic history, birth-related measurement, and genetics.

Design, Setting, and Participants

This study was conducted from May 1, 1981, to December 31, 2016. For the analysis, which used a Danish population-based case-cohort sample of individuals born between 1981 and 2005, genotype data and matched longitudinal health register data were taken from the longitudinal Danish population-based Integrative Psychiatric Research Consortium 2012 case-cohort study. Included were individuals with mental disorders (attention-deficit/hyperactivity disorder [ADHD]), autism spectrum disorder (ASD), major depressive disorder (MDD), bipolar disorder (BD), schizophrenia spectrum disorders (SCZ), and population controls. Data were analyzed from February 1, 2021, to January 24, 2022.

Exposure

At least 1 hospital contact with diagnosis of ADHD, ASD, MDD, BD, or SCZ.

Main Outcomes and Measures

The predictability of (1) mental disorder diagnosis and (2) severity trajectories (measured by future outpatient hospital contacts, admissions, and suicide attempts) were investigated using both a cross-diagnostic and single-disorder setup. Predictive power was measured by AUC, accuracy, and Matthews correlation coefficient (MCC), including an estimate of feature importance.

Results

A total of 63 535 individuals (mean [SD] age, 23 [7] years; 34 944 male [55%]; 28 591 female [45%]) were included in the model. Based on data prior to diagnosis, the specific diagnosis was predicted in a multidiagnostic prediction model including the background population with an overall area under the curve (AUC) of 0.81 and MCC of 0.28, whereas the single-disorder models gave AUCs/MCCs of 0.84/0.54 for SCZ, 0.79/0.41 for BD, 0.77/0.39 for ASD, 0.74/0.38, for ADHD, and 0.74/0.38 for MDD. The most important data sets for multidiagnostic prediction were previous mental disorders and age (11%-23% reduction in prediction accuracy when removed) followed by family diagnoses, birth-related measurements, and genetic data (3%-5% reduction in prediction accuracy when removed). Furthermore, when predicting subsequent disease trajectories of the disorder, the most severe cases were the most easily predictable, with an AUC of 0.72.

Conclusions and Relevance

Results of this diagnostic study suggest the possibility of combining genetics and registry data to predict both mental disorder diagnosis and disorder progression in a clinically relevant, cross-diagnostic setting prior to clinical assessment.

Machine Learning-Based Mortality Prediction Model for Critically Ill Cancer Patients Admitted to the Intensive Care Unit (CanICU)

BACKGROUND

Although cancer patients are increasingly admitted to the intensive care unit (ICU) for cancer- or treatment-related complications, improved mortality prediction remains a big challenge. This study describes a new ML-based mortality prediction model for critically ill cancer patients admitted to ICU.

PATIENTS AND METHODS

We developed CanICU, a machine learning-based 28-day mortality prediction model for adult cancer patients admitted to ICU from Medical Information Mart for Intensive Care (MIMIC) database in the USA (n = 766), Yonsei Cancer Center (YCC, n = 3571), and Samsung Medical Center in Korea (SMC, n = 2563) from 2 January 2008 to 31 December 2017. The accuracy of CanICU was measured using sensitivity, specificity, and area under the receiver operating curve (AUROC).

RESULTS

A total of 6900 patients were included, with a 28-day mortality of 10.2%/12.7%/36.6% and a 1-year mortality of 30.0%/36.6%/58.5% in the YCC, SMC, and MIMIC-III cohort. Nine clinical and laboratory factors were used to construct the classifier using a random forest machine-learning algorithm. CanICU had 96% sensitivity/73% specificity with the area under the receiver operating characteristic (AUROC) of 0.94 for 28-day, showing better performance than current prognostic models, including the Acute Physiology and Chronic Health Evaluation (APACHE) or Sequential Organ Failure Assessment (SOFA) score. Application of CanICU in two external data sets across the countries yielded 79-89% sensitivity, 58-59% specificity, and 0.75-0.78 AUROC for 28-day mortality. The CanICU score was also correlated with one-year mortality with 88-93% specificity.

CONCLUSION

CanICU offers improved performance for predicting mortality in critically ill cancer patients admitted to ICU. A user-friendly online implementation is available and should be valuable for better mortality risk stratification to allocate ICU care for cancer patients.

Clinical utility of automatic phenotype annotation in unstructured clinical notes: intensive care unit use

OBJECTIVE

Clinical notes contain information that has not been documented elsewhere, including responses to treatment and clinical findings, which are crucial for predicting key outcomes in patients in acute care. In this study, we propose the automatic annotation of phenotypes from clinical notes as a method to capture essential information to predict outcomes in the intensive care unit (ICU). This information is complementary to typically used vital signs and laboratory test results.

METHODS

In this study, we developed a novel phenotype annotation model to extract the phenotypical features of patients, which were then used as input features of predictive models to predict ICU patient outcomes. We demonstrated and validated this approach by conducting experiments on three ICU prediction tasks, including in-hospital mortality, physiological decompensation and length of stay (LOS) for over 24 000 patients using the Medical Information Mart for Intensive Care (MIMIC-III) dataset.

RESULTS

The predictive models incorporating phenotypical information achieved 0.845 (area under the curve-receiver operating characteristic (AUC-ROC)) for in-hospital mortality, 0.839 (AUC-ROC) for physiological decompensation and 0.430 (kappa) for LOS, all of which consistently outperformed the baseline models using only vital signs and laboratory test results. Moreover, we conducted a thorough interpretability study showing that phenotypes provide valuable insights at both the patient and cohort levels.

CONCLUSION

The proposed approach demonstrates that phenotypical information complements traditionally used vital signs and laboratory test results and significantly improves the accuracy of outcome prediction in the ICU.

Explainable Preoperative Automated Machine Learning Prediction Model for Cardiac Surgery-Associated Acute Kidney Injury

BACKGROUND

We aimed to develop and validate an automated machine learning (autoML) prediction model for cardiac surgery-associated acute kidney injury (CSA-AKI).

METHODS

Using 69 preoperative variables, we developed several models to predict post-operative AKI in adult patients undergoing cardiac surgery. Models included autoML and non-autoML types, including decision tree (DT), random forest (RF), extreme gradient boosting (XGBoost), and artificial neural network (ANN), as well as a logistic regression prediction model. We then compared model performance using area under the receiver operating characteristic curve (AUROC) and assessed model calibration using Brier score on the independent testing dataset.

RESULTS

The incidence of CSA-AKI was 36%. Stacked ensemble autoML had the highest predictive performance among autoML models, and was chosen for comparison with other non-autoML and multivariable logistic regression models. The autoML had the highest AUROC (0.79), followed by RF (0.78), XGBoost (0.77), multivariable logistic regression (0.77), ANN (0.75), and DT (0.64). The autoML had comparable AUROC with RF and outperformed the other models. The autoML was well-calibrated. The Brier score for autoML, RF, DT, XGBoost, ANN, and multivariable logistic regression was 0.18, 0.18, 0.21, 0.19, 0.19, and 0.18, respectively. We applied SHAP and LIME algorithms to our autoML prediction model to extract an explanation of the variables that drive patient-specific predictions of CSA-AKI.

CONCLUSION

We were able to present a preoperative autoML prediction model for CSA-AKI that provided high predictive performance that was comparable to RF and superior to other ML and multivariable logistic regression models. The novel approaches of the proposed explainable preoperative autoML prediction model for CSA-AKI may guide clinicians in advancing individualized medicine plans for patients under cardiac surgery.

Discrete-time survival analysis in the critically ill: a deep learning approach using heterogeneous data

Prediction of survival for patients in intensive care units (ICUs) has been subject to intense research. However, no models exist that embrace the multiverse of data in ICUs. It is an open question whether deep learning methods using automated data integration with minimal pre-processing of mixed data domains such as free text, medical history and high-frequency data can provide discrete-time survival estimates for individual ICU patients. We trained a deep learning model on data from patients admitted to ten ICUs in the Capital Region of Denmark and the Region of Southern Denmark between 2011 and 2018. Inspired by natural language processing we mapped the electronic patient record data to an embedded representation and fed the data to a recurrent neural network with a multi-label output layer representing the chance of survival at different follow-up times. We evaluated the performance using the time-dependent concordance index. In addition, we quantified and visualized the drivers of survival predictions using the SHAP methodology. We included 37,355 admissions of 29,417 patients in our study. Our deep learning models outperformed traditional Cox proportional-hazard models with concordance index in the ranges 0.72–0.73, 0.71–0.72, 0.71, and 0.69–0.70, for models applied at baseline 0, 24, 48, and 72 h, respectively. Deep learning models based on a combination of entity embeddings and survival modelling is a feasible approach to obtain individualized survival estimates in data-rich settings such as the ICU. The interpretable nature of the models enables us to understand the impact of the different data domains.

ConSIG: consistent discovery of molecular signature from OMIC data

The discovery of proper molecular signature from OMIC data is indispensable for determining biological state, physiological condition, disease etiology, and therapeutic response. However, the identified signature is reported to be highly inconsistent, and there is little overlap among the signatures identified from different biological datasets. Such inconsistency raises doubts about the reliability of reported signatures and significantly hampers its biological and clinical applications. Herein, an online tool, ConSIG, was constructed to realize consistent discovery of gene/protein signature from any uploaded transcriptomic/proteomic data. This tool is unique in a) integrating a novel strategy capable of significantly enhancing the consistency of signature discovery, b) determining the optimal signature by collective assessment, and c) confirming the biological relevance by enriching the disease/gene ontology. With the increasingly accumulated concerns about signature consistency and biological relevance, this online tool is expected to be used as an essential complement to other existing tools for OMIC-based signature discovery. ConSIG is freely accessible to all users without login requirement at https://idrblab.org/consig/.

Training prediction models for individual risk assessment of postoperative complications after surgery for colorectal cancer

BACKGROUND

The occurrence of postoperative complications and anastomotic leakage are major drivers of mortality in the immediate phase after colorectal cancer surgery. We trained prediction models for calculating patients' individual risk of complications based only on preoperatively available data in a multidisciplinary team setting. Knowing prior to surgery the probability of developing a complication could aid in improving informed decision-making by surgeon and patient and individualize surgical treatment trajectories.

METHODS

All patients over 18 years of age undergoing any resection for colorectal cancer between January 1, 2014 and December 31, 2019 from the nationwide Danish Colorectal Cancer Group database were included. Data from the database were converted into Observational Medical Outcomes Partnership Common Data Model maintained by the Observation Health Data Science and Informatics initiative. Multiple machine learning models were trained to predict postoperative complications of Clavien-Dindo grade ≥ 3B and anastomotic leakage within 30 days after surgery.

RESULTS

Between 2014 and 2019, 23,907 patients underwent resection for colorectal cancer in Denmark. A Clavien-Dindo complication grade ≥ 3B occurred in 2,958 patients (12.4%). Of 17,190 patients that received an anastomosis, 929 experienced anastomotic leakage (5.4%). Among the compared machine learning models, Lasso Logistic Regression performed best. The predictive model for complications had an area under the receiver operating characteristic curve (AUROC) of 0.704 (95%CI 0.683-0.724) and an AUROC of 0.690 (95%CI 0.655-0.724) for anastomotic leakage.

CONCLUSIONS

The prediction of postoperative complications based only on preoperative variables using a national quality assurance colorectal cancer database shows promise for calculating patient's individual risk. Future work will focus on assessing the value of adding laboratory parameters and drug exposure as candidate predictors. Furthermore, we plan to assess the external validity of our proposed model.

Osteoarthritis endotype discovery via clustering of biochemical marker data

OBJECTIVES

Osteoarthritis (OA) patient stratification is an important challenge to design tailored treatments and drive drug development. Biochemical markers reflecting joint tissue turnover were measured in the IMI-APPROACH cohort at baseline and analysed using a machine learning approach in order to study OA-dominant phenotypes driven by the endotype-related clusters and discover the driving features and their disease-context meaning.

METHOD

Data quality assessment was performed to design appropriate data preprocessing techniques. The k-means clustering algorithm was used to find dominant subgroups of patients based on the biochemical markers data. Classification models were trained to predict cluster membership, and Explainable AI techniques were used to interpret these to reveal the driving factors behind each cluster and identify phenotypes. Statistical analysis was performed to compare differences between clusters with respect to other markers in the IMI-APPROACH cohort and the longitudinal disease progression.

RESULTS

Three dominant endotypes were found, associated with three phenotypes: C1) low tissue turnover (low repair and articular cartilage/subchondral bone turnover), C2) structural damage (high bone formation/resorption, cartilage degradation) and C3) systemic inflammation (joint tissue degradation, inflammation, cartilage degradation). The method achieved consistent results in the FNIH/OAI cohort. C1 had the highest proportion of non-progressors. C2 was mostly linked to longitudinal structural progression, and C3 was linked to sustained or progressive pain.

CONCLUSIONS

This work supports the existence of differential phenotypes in OA. The biomarker approach could potentially drive stratification for OA clinical trials and contribute to precision medicine strategies for OA progression in the future.

TRIAL REGISTRATION NUMBER

NCT03883568.

Machine Learning Prediction Models for Mortality in Intensive Care Unit Patients with Lactic Acidosis

Background: Lactic acidosis is the most common cause of anion gap metabolic acidosis in the intensive care unit (ICU), associated with poor outcomes including mortality. We sought to compare machine learning (ML) approaches versus logistic regression analysis for prediction of mortality in lactic acidosis patients admitted to the ICU. Methods: We used the Medical Information Mart for Intensive Care (MIMIC-III) database to identify ICU adult patients with lactic acidosis (serum lactate ≥4 mmol/L). The outcome of interest was hospital mortality. We developed prediction models using four ML approaches consisting of random forest (RF), decision tree (DT), extreme gradient boosting (XGBoost), artificial neural network (ANN), and statistical modeling with forward stepwise logistic regression using the testing dataset. We then assessed model performance using area under the receiver operating characteristic curve (AUROC), accuracy, precision, error rate, Matthews correlation coefficient (MCC), F1 score, and assessed model calibration using the Brier score, in the independent testing dataset. Results: Of 1919 lactic acidosis ICU patients, 1535 and 384 were included in the training and testing dataset, respectively. Hospital mortality was 30%. RF had the highest AUROC at 0.83, followed by logistic regression 0.81, XGBoost 0.81, ANN 0.79, and DT 0.71. In addition, RF also had the highest accuracy (0.79), MCC (0.45), F1 score (0.56), and lowest error rate (21.4%). The RF model was the most well-calibrated. The Brier score for RF, DT, XGBoost, ANN, and multivariable logistic regression was 0.15, 0.19, 0.18, 0.19, and 0.16, respectively. The RF model outperformed multivariable logistic regression model, SOFA score (AUROC 0.74), SAP II score (AUROC 0.77), and Charlson score (AUROC 0.69). Conclusion: The ML prediction model using RF algorithm provided the highest predictive performance for hospital mortality among ICU patient with lactic acidosis.

Optimizing drug selection from a prescription trajectory of one patient

It is unknown how sequential drug patterns convey information on a patient's health status and treatment guidelines rarely account for this. Drug-agnostic longitudinal analyses of prescription trajectories in a population-wide setting are needed. In this cohort study, we used 24 years of data (1.1 billion prescriptions) from the Danish prescription registry to model the risk of sequentially redeeming a drug after another. Drug pairs were used to build multistep longitudinal prescription trajectories. These were subsequently used to stratify patients and calculate survival hazard ratios between the stratified groups. The similarity between prescription histories was used to determine individuals' best treatment option. Over the course of 122 million person-years of observation, we identified 9 million common prescription trajectories and demonstrated their predictive power using hypertension as a case. Among patients treated with agents acting on the renin-angiotensin system we identified four groups: patients prescribed angiotensin converting enzyme (ACE) inhibitor without change, angiotensin receptor blockers (ARBs) without change, ACE with posterior change to ARB, and ARB posteriorly changed to ACE. In an adjusted time-to-event analysis, individuals treated with ACE compared to those treated with ARB had lower survival probability (hazard ratio, 0.73 [95% CI, 0.64-0.82]; P < 1 × 10^-16). Replication in UK Biobank data showed the same trends. Prescription trajectories can provide novel insights into how individuals' drug use change over time, identify suboptimal or futile prescriptions and suggest initial treatments different from first line therapies. Observations of this kind may also be important when updating treatment guidelines.

Comparison of a polypharmacy-based scale with Charlson comorbidity index to predict 6-month mortality in chronic complex patients after an ED visit

AIMS

The aim of this study was to test whether a newly designed polypharmacy-based scale would perform better than Charlson's Comorbidity Index (CCI) to predict outcomes in chronic complex adult patients after a reference Emergency Department (ED) visit.

METHODS

We built a polypharmacy-based scale with prespecified drug families. The primary outcome was 6-month mortality after the reference ED visit. Predefined secondary outcomes were need for hospital admission, 30-day readmission, and 30-day and 90-day mortality. We evaluated the ability of the CCI and the polypharmacy-based scale to independently predict 6-month mortality using logistic regression, receiver operating characteristic (ROC) curves, and cumulative survival curves using Kaplan-Meier estimates and the log-rank test for three-category distributions of the polypharmacy-based scale and the CCI. Finally, we sought to replicate our results in two different external validation cohorts.

RESULTS

We included 201 patients (53.7% women, mean age = 81.4 years), 162 of whom were admitted to the hospital at the reference ED visit. In separate multivariable analyses accounting for gender, age and main diagnosis at discharge, both the polypharmacy-based scale (P < .001) and the CCI (P = .005) independently predicted 6-month mortality. The polypharmacy-based scale performed better in the ROC analyses (area under the curve [AUC] = 0.838, 95% confidence interval [CI] = 0.780-0.896) than the CCI (AUC = 0.628, 95% CI = 0.548-0.707). In the 6-month cumulative survival analysis, the polypharmacy-based scale showed statistical significance (P < .001), whereas the CCI did not (P = .484). We replicated our results in the validation cohorts.

CONCLUSIONS

Our polypharmacy-based scale performed significantly better than the CCI to predict 6-month mortality in chronic complex patients after a reference ED visit.

The management of multi-morbidity in elderly patients: Ready yet for precision medicine in intensive care?

There is ongoing demographic ageing and increasing longevity of the population, with previously devastating and often-fatal diseases now transformed into chronic conditions. This is turning multi-morbidity into a major challenge in the world of critical care. After many years of research and innovation, mainly in geriatric care, the concept of multi-morbidity now requires fine-tuning to support decision-making for patients along their whole trajectory in healthcare, including in the intensive care unit (ICU). This article will discuss current challenges and present approaches to adapt critical care services to the needs of these patients.

A catalogue of tools and variables from crisis and routine care to support decision-making about allocation of intensive care beds and ventilator treatment during pandemics: Scoping review

PURPOSE

This scoping review sought to identify objective factors to assist clinicians and policy-makers in making consistent, objective and ethically sound decisions about resource allocation when healthcare rationing is inevitable.

MATERIALS AND METHODS

Review of guidelines and tools used in ICUs, hospital wards and emergency departments on how to best allocate intensive care beds and ventilators either during routine care or developed during previous epidemics, and association with patient outcomes during and after hospitalisation.

RESULTS

Eighty publications from 20 countries reporting accuracy or validity of prognostic tools/algorithms, or significant correlation between prognostic variables and clinical outcomes met our eligibility criteria: twelve pandemic guidelines/triage protocols/consensus statements, twenty-two pandemic algorithms, and 46 prognostic tools/variables from non-crisis situations. Prognostic indicators presented here can be combined to create locally-relevant triage algorithms for clinicians and policy makers deciding about allocation of ICU beds and ventilators during a pandemic. No consensus was found on the ethical issues to incorporate in the decision to admit or triage out of intensive care.

CONCLUSIONS

This review provides a unique reference intended as a discussion starter for clinicians and policy makers to consider formalising an objective a locally-relevant triage consensus document that enhances confidence in decision-making during healthcare rationing of critical care and ventilator resources.

Evaluation of a hypothetical decision-support tool for intensive care triage of patients with coronavirus disease 2019 (COVID-19)

Background: At the start of the coronavirus disease 2019 (COVID-19) pandemic there was widespread concern about potentially overwhelming demand for intensive care and the need for intensive care unit (ICU) triage. In March 2020, a draft United Kingdom (UK) guideline proposed a decision-support tool (DST). We sought to evaluate the accuracy of the tool in patients with COVID-19. Methods: We retrospectively identified patients in two groups: referred and not referred to intensive care in a single UK national health service (NHS) trust in April 2020. Age, Clinical Frailty Scale score (CFS), and co-morbidities were collected from patients’ records and recorded, along with ceilings of treatment and outcome. We compared the DST, CFS, and age alone as predictors of mortality, and treatment ceiling decisions. Results: In total, 151 patients were included in the analysis, with 75 in the ICU and 76 in the non-ICU-reviewed groups. Age, clinical frailty and DST score were each associated with increased mortality and higher likelihood of treatment limitation (p-values all <.001). A DST cut-off score of >8 had 65% (95% confidence interval (CI) 51%-79%) sensitivity and 63% (95% CI 54%-72%) specificity for predicting mortality. It had a sensitivity of 80% (70%-88%) and specificity of 96% (95% CI 90%-100%) for predicting treatment limitation. The DST was more discriminative than age alone (p<0.001), and potentially more discriminative than CFS (p=0.08) for predicting treatment ceiling decisions. Conclusions: During the first wave of the COVID-19 pandemic, in a hospital without severe resource limitations, a hypothetical decision support tool was limited in its predictive value for mortality, but appeared to be sensitive and specific for predicting treatment limitation.

Non-occlusive mesenteric ischemia: Diagnostic challenges and perspectives in the era of artificial intelligence

Acute mesenteric ischemia (AMI) is a severe condition associated with poor prognosis, ultimately leading to death due to multiorgan failure. Several mechanisms may lead to AMI, and non-occlusive mesenteric ischemia (NOMI) represents a particular form of AMI. NOMI is prevalent in intensive care units in critically ill patients. In NOMI management, promptness and accuracy of diagnosis are paramount to achieve decisive treatment, but the last decades have been marked by failure to improve NOMI prognosis, due to lack of tools to detect this condition. While real-life diagnostic management relies on a combination of physical examination, several biomarkers, imaging, and endoscopy to detect the possibility of several grades of NOMI, research studies only focus on a few elements at a time. In the era of artificial intelligence (AI), which can aggregate thousands of variables in complex longitudinal models, the prospect of achieving accurate diagnosis through machine-learning-based algorithms may be sought. In the following work, we bring you a state-of-the-art literature review regarding NOMI, its presentation, its mechanics, and the pitfalls of routine work-up diagnostic exams including biomarkers, imaging, and endoscopy, we raise the perspectives of new biomarker exams, and finally we discuss what AI may add to the field, after summarizing what this technique encompasses.

Long term outcomes following critical care hospital admission: A prospective cohort study of UK biobank participants✰,★

BACKGROUND

: This study aimed to understand the impact of a critical care admission on long-term outcomes, compared to other hospitalised patients without a critical care encounter. A secondary aim was to examine the interrelationship between emotional, physical, and social problems during recovery.

METHODS

: We utilised data from the UK Biobank, an on-going, prospective population-based cohort study. We employed propensity score matching to assess differences in outcomes between patients with a critical care encounter and patients admitted to the hospital (first admission to hospital available) without critical care. Structural equation modelling was used to analyse emotional, physical and social outcomes following critical illness and the relationships between these health domains.

FINDINGS

: Data from 1,618 patients were analysed. The median time to follow-up in the critical care cohort was 4427 days (IQR:788-6146) vs 4516 days (IQR: 811-6369) in the non-critical care, hospitalised cohort. Across the two time periods assessed (pre and post 2000), patients exposed to critical care were more likely to experience mental health issues such as depression (p < 0.01) and social isolation (p = 0.01) following discharge from hospital. The critical care cohort were also more likely to have social problems such as the requirement for government funded welfare support (p = 0.02). In the critical care cohort, social and emotional health were closely correlated (p < 0.001, 95% CI:0.33-0.54). The nature of physical problems changed over time; pre-2000 there was a significant difference between the critical and non-critical care in physical outcomes following discharge from hospital, however, there was no difference detected between the two cohorts post-2000.

INTERPRETATION

This cohort study has demonstrated that survivors of critical illness have different psycho-social outcomes to matched patients, hospitalised without a critical care encounter.

FUNDING

JM is funded by a THIS.Institute (University of Cambridge) Research Fellowship (PD-2019-02-16). AHL is part of the Social and Public Health Sciences Unit, funded by the Medical Research Council (MC_UU_12017/13) and the Scottish Government Chief Scientist Office (SPHSU13).

Early Prediction of Mortality, Severity, and Length of Stay in the Intensive Care Unit of Sepsis Patients Based on Sepsis 3.0 by Machine Learning Models

Background: Early prediction of the clinical outcome of patients with sepsis is of great significance and can guide treatment and reduce the mortality of patients. However, it is clinically difficult for clinicians. Methods: A total of 2,224 patients with sepsis were involved over a 3-year period (2016-2018) in the intensive care unit (ICU) of Peking Union Medical College Hospital. With all the key medical data from the first 6 h in the ICU, three machine learning models, logistic regression, random forest, and XGBoost, were used to predict mortality, severity (sepsis/septic shock), and length of ICU stay (LOS) (>6 days, ≤ 6 days). Missing data imputation and oversampling were completed on the dataset before introduction into the models. Results: Compared to the mortality and LOS predictions, the severity prediction achieved the best classification results, based on the area under the operating receiver characteristics (AUC), with the random forest classifier (sensitivity = 0.65, specificity = 0.73, F1 score = 0.72, AUC = 0.79). The random forest model also showed the best overall performance (mortality prediction: sensitivity = 0.50, specificity = 0.84, F1 score = 0.66, AUC = 0.74; LOS prediction: sensitivity = 0.79, specificity = 0.66, F1 score = 0.69, AUC = 0.76) among the three models. The predictive ability of the SOFA score itself was inferior to that of the above three models. Conclusions: Using the random forest classifier in the first 6 h of ICU admission can provide a comprehensive early warning of sepsis, which will contribute to the formulation and management of clinical decisions and the allocation and management of resources.

Identifying and characterizing high-risk clusters in a heterogeneous ICU population with deep embedded clustering

Critically ill patients constitute a highly heterogeneous population, with seemingly distinct patients having similar outcomes, and patients with the same admission diagnosis having opposite clinical trajectories. We aimed to develop a machine learning methodology that identifies and provides better characterization of patient clusters at high risk of mortality and kidney injury. We analysed prospectively collected data including co-morbidities, clinical examination, and laboratory parameters from a minimally-selected population of 743 patients admitted to the ICU of a Dutch hospital between 2015 and 2017. We compared four clustering methodologies and trained a classifier to predict and validate cluster membership. The contribution of different variables to the predicted cluster membership was assessed using SHapley Additive exPlanations values. We found that deep embedded clustering yielded better results compared to the traditional clustering algorithms. The best cluster configuration was achieved for 6 clusters. All clusters were clinically recognizable, and differed in in-ICU, 30-day, and 90-day mortality, as well as incidence of acute kidney injury. We identified two high mortality risk clusters with at least 60%, 40%, and 30% increased. ICU, 30-day and 90-day mortality, and a low risk cluster with 25–56% lower mortality risk. This machine learning methodology combining deep embedded clustering and variable importance analysis, which we made publicly available, is a possible solution to challenges previously encountered by clustering analyses in heterogeneous patient populations and may help improve the characterization of risk groups in critical care.

Use of deep learning to develop continuous-risk models for adverse event prediction from electronic health records

Early prediction of patient outcomes is important for targeting preventive care. This protocol describes a practical workflow for developing deep-learning risk models that can predict various clinical and operational outcomes from structured electronic health record (EHR) data. The protocol comprises five main stages: formal problem definition, data pre-processing, architecture selection, calibration and uncertainty, and generalizability evaluation. We have applied the workflow to four endpoints (acute kidney injury, mortality, length of stay and 30-day hospital readmission). The workflow can enable continuous (e.g., triggered every 6 h) and static (e.g., triggered at 24 h after admission) predictions. We also provide an open-source codebase that illustrates some key principles in EHR modeling. This protocol can be used by interdisciplinary teams with programming and clinical expertise to build deep-learning prediction models with alternate data sources and prediction tasks.

Chance of live birth: a nationwide, registry-based cohort study

STUDY QUESTION

Does the sequence of prior pregnancy events (pregnancy losses, live births, ectopic pregnancies, molar pregnancy and still birth), obstetric complications and maternal age affect chance of live birth in the next pregnancy and are prior events predictive for the outcome?

SUMMARY ANSWER

The sequence of pregnancy outcomes is significantly associated with chance of live birth; however, pregnancy history and age are insufficient to predict the outcome of an individual woman's next pregnancy.

WHAT IS KNOWN ALREADY

Adverse pregnancy outcomes decrease the chance of live birth in the next pregnancy, whereas the impact of prior live births is less clear.

STUDY DESIGN, SIZE, DURATION

Nationwide, registry-based cohort study of 1 285 230 women with a total of 2 722 441 pregnancies from 1977 to 2017.

PARTICIPANTS/MATERIALS, SETTING, METHODS

All women living in Denmark in the study period with at least one pregnancy in either the Danish Medical Birth Registry or the Danish National Patient Registry. Data were analysed using logistic regression with a robust covariance model to account for women with more than one pregnancy. Model discrimination and calibration were ascertained using 20% of the women in the cohort randomly selected as an internal validation set.

MAIN RESULTS AND THE ROLE OF CHANCE

Obstetric complications, still birth, ectopic pregnancies and pregnancy losses had a negative effect on the chance of live birth in the next pregnancy. Consecutive, identical pregnancy outcomes (pregnancy losses, live births or ectopic pregnancies) immediately preceding the next pregnancy had a larger impact than the total number of any outcome. Model discrimination was modest (C-index = 0.60, positive predictive value = 0.45), but the models were well calibrated.

LIMITATIONS, REASONS FOR CAUTION

While prior pregnancy outcomes and their sequence significantly influenced the chance of live birth, the discriminative abilities of the predictive models demonstrate clearly that pregnancy history and maternal age are insufficient to reliably predict the outcome of a given pregnancy.

WIDER IMPLICATIONS OF THE FINDINGS

Prior pregnancy history has a significant impact on the chance of live birth in the next pregnancy. However, the results emphasize that only taking age and number of losses into account does not predict if a pregnancy will end as a live birth or not. A better understanding of biological determinants for pregnancy outcomes is urgently needed.

STUDY FUNDING/COMPETING INTEREST(S)

The work was supported by the Novo Nordisk Foundation, Ole Kirk Foundation and Rigshospitalet's Research Foundation. The authors have no financial relationships that could appear to have influenced the work.

TRIAL REGISTRATION NUMBER

N/A.

Prediction of 90-day mortality after surgery for colorectal cancer using standardized nationwide quality-assurance data

Background

Personalized risk assessment provides opportunities for tailoring treatment, optimizing healthcare resources and improving outcome. The aim of this study was to develop a 90-day mortality-risk prediction model for identification of high- and low-risk patients undergoing surgery for colorectal cancer.

Methods

This was a nationwide cohort study using records from the Danish Colorectal Cancer Group database that included all patients undergoing surgery for colorectal cancer between 1 January 2004 and 31 December 2015. A least absolute shrinkage and selection operator logistic regression prediction model was developed using 121 pre- and intraoperative variables and internally validated in a hold-out test data set. The accuracy of the model was assessed in terms of discrimination and calibration.

Results

In total, 49 607 patients were registered in the database. After exclusion of 16 680 individuals, 32 927 patients were included in the analysis. Overall, 1754 (5.3 per cent) deaths were recorded. Targeting high-risk individuals, the model identified 5.5 per cent of all patients facing a risk of 90-day mortality exceeding 35 per cent, corresponding to a 6.7 times greater risk than the average population. Targeting low-risk individuals, the model identified 20.9 per cent of patients facing a risk less than 0.3 per cent, corresponding to a 17.7 times lower risk compared with the average population. The model exhibited discriminatory power with an area under the receiver operating characteristics curve of 85.3 per cent (95 per cent c.i. 83.6 to 87.0) and excellent calibration with a Brier score of 0.04 and 32 per cent average precision.

Conclusion

Pre- and intraoperative data, as captured in national health registries, can be used to predict 90-day mortality accurately after colorectal cancer surgery.

'Leading by Science' through Covid-19: the NHS Data Store & Automated Decision-Making

The UK government announced in March 2020 that it would create an NHS Covid-19 ‘Data Store’ from information routinely collected as part of the health service. This ‘Store’ would use a number of sources of population data to provide a ‘single source of truth’ about the spread of the coronavirus in England. The initiative illustrates the difficulty of relying on automated processing when making healthcare decisions under the General Data Protection Regulation (GDPR). The end-product of the store, a number of ‘dashboards’ for decision-makers, was intended to include models and simulations developed through artificial intelligence. Decisions made on the basis of these dashboards would be significant, even (it was suggested) to the point of diverting patients and critical resources between hospitals based on their predictions.

How these models will be developed, and externally validated, remains unclear. This is an issue if they are intended to be used for decisions which will affect patients so directly and acutely. We have (by default) a right under the GDPR not to be subject to significant decisions based solely on automated decision-making. It is not obvious, at present, whether resource allocation within the NHS could take place in reliance on this automated modelling. The recent A Level debacle illustrates, in the context of education, the risks of basing life-changing decisions on the national application of a single equation. It is worth considering the potential consequences for the health service if the NHS Data Store is used for resource planning as part of the Covid-19 response.

Theorising survivorship after intensive care: A systematic review of patient and family experiences

AIMS AND OBJECTIVE

This systematic literature review explores and maps what we know about survivorship to understand how survivorship can be theoretically defined.

BACKGROUND

Survivorship of critical illness has been identified as a challenge for the 21st Century. Whilst the use of the term 'survivorship' is now common in critical care, it has been borrowed from the cancer literature where the discourse on what survivorship means in a cancer context is ongoing and remains largely descriptive. In the absence of a theoretical understanding, the term 'survivorship' is often used in critical illness in a generic way, limiting our understanding of what survivorship is. The current COVID-19 pandemic adds to an urgency of understanding what intensive care unit (ICU) survivorship might mean, given the emerging long-term consequences of this patient cohort. We set out to explore how survivorship after critical illness is being conceptualised and what the implications might be for clinical practice and research.

DESIGN

Integrated systematic literature review. The review protocol was registered with PROSPERO International Prospective Register of Systematic Reviews. PRISMA guidelines were followed and a PRISMA checklist for reporting systematic reviews completed.

RESULTS

The three main themes around which the reviewed studies were organised are: (a) healthcare system; (b) ICU survivors' families; and (c) ICU survivor's identity. These three themes feed into an overarching core theme of 'ICU Survivorship Experiences'. These themes map our current knowledge of what happens when a patient survives a critical illness and where we are in understanding ICU survivorship.

CONCLUSION

We mapped in this systematic review the different pieces of the jigsaw that emerge following critical illness to understand and see the bigger picture of what happens after patients survive critical illness. It is evident that existing research has mapped these connections, but what we have not managed to do yet is defining what survivorship is theoretically. We offer a preliminary definition of survivorship as a process but are aware that this definition needs to be developed further with patients and families.

Noninvasive Real-Time Mortality Prediction in Intensive Care Units Based on Gradient Boosting Method: Model Development and Validation Study

BACKGROUND

Monitoring critically ill patients in intensive care units (ICUs) in real time is vitally important. Although scoring systems are most often used in risk prediction of mortality, they are usually not highly precise, and the clinical data are often simply weighted. This method is inefficient and time-consuming in the clinical setting.

OBJECTIVE

The objective of this study was to integrate all medical data and noninvasively predict the real-time mortality of ICU patients using a gradient boosting method. Specifically, our goal was to predict mortality using a noninvasive method to minimize the discomfort to patients.

METHODS

In this study, we established five models to predict mortality in real time based on different features. According to the monitoring, laboratory, and scoring data, we constructed the feature engineering. The five real-time mortality prediction models were RMM (based on monitoring features), RMA (based on monitoring features and the Acute Physiology and Chronic Health Evaluation [APACHE]), RMS (based on monitoring features and Sequential Organ Failure Assessment [SOFA]), RMML (based on monitoring and laboratory features), and RM (based on all monitoring, laboratory, and scoring features). All models were built using LightGBM and tested with XGBoost. We then compared the performance of all models, with particular focus on the noninvasive method, the RMM model.

RESULTS

After extensive experiments, the area under the curve of the RMM model was 0.8264, which was superior to that of the RMA and RMS models. Therefore, predicting mortality using the noninvasive method was both efficient and practical, as it eliminated the need for extra physical interventions on patients, such as the drawing of blood. In addition, we explored the top nine features relevant to real-time mortality prediction: invasive mean blood pressure, heart rate, invasive systolic blood pressure, oxygen concentration, oxygen saturation, balance of input and output, total input, invasive diastolic blood pressure, and noninvasive mean blood pressure. These nine features should be given more focus in routine clinical practice.

CONCLUSIONS

The results of this study may be helpful in real-time mortality prediction in patients in the ICU, especially the noninvasive method. It is efficient and favorable to patients, which offers a strong practical significance.

A machine learning approach to predict healthcare-associated infections at intensive care unit admission: findings from the SPIN-UTI project

BACKGROUND

Identifying patients at higher risk of healthcare-associated infections (HAIs) in intensive care units (ICUs) represents a major challenge for public health. Machine learning could improve patient risk stratification and lead to targeted infection prevention and control interventions.

AIM

To evaluate the performance of the Simplified Acute Physiology Score (SAPS) II for HAI risk prediction in ICUs, using both traditional statistical and machine learning approaches.

METHODS

Data for 7827 patients from the 'Italian Nosocomial Infections Surveillance in Intensive Care Units' project were used in this study. The Support Vector Machines (SVM) algorithm was applied to classify patients according to sex, patient origin, non-surgical treatment for acute coronary disease, surgical intervention, SAPS II at admission, presence of invasive devices, trauma, impaired immunity, and antibiotic therapy in 48 h preceding ICU admission.

FINDINGS

The performance of SAPS II for predicting HAI risk provides a receiver operating characteristic curve with an area under the curve of 0.612 (P<0.001) and accuracy of 56%. Considering SAPS II along with other characteristics at ICU admission, the SVM classifier was found to have accuracy of 88% and an AUC of 0.90 (P<0.001) for the test set. The predictive ability was lower when considering the same SVM model but with the SAPS II variable removed (accuracy 78%, AUC 0.66).

CONCLUSIONS

This study suggested that the SVM model is a useful tool for early prediction of patients at higher risk of HAIs at ICU admission.

Early Prediction of Seven-Day Mortality in Intensive Care Unit Using a Machine Learning Model: Results from the SPIN-UTI Project

Patients in intensive care units (ICUs) were at higher risk of worsen prognosis and mortality. Here, we aimed to evaluate the ability of the Simplified Acute Physiology Score (SAPS II) to predict the risk of 7-day mortality, and to test a machine learning algorithm which combines the SAPS II with additional patients' characteristics at ICU admission. We used data from the "Italian Nosocomial Infections Surveillance in Intensive Care Units" network. Support Vector Machines (SVM) algorithm was used to classify 3782 patients according to sex, patient's origin, type of ICU admission, non-surgical treatment for acute coronary disease, surgical intervention, SAPS II, presence of invasive devices, trauma, impaired immunity, antibiotic therapy and onset of HAI. The accuracy of SAPS II for predicting patients who died from those who did not was 69.3%, with an Area Under the Curve (AUC) of 0.678. Using the SVM algorithm, instead, we achieved an accuracy of 83.5% and AUC of 0.896. Notably, SAPS II was the variable that weighted more on the model and its removal resulted in an AUC of 0.653 and an accuracy of 68.4%. Overall, these findings suggest the present SVM model as a useful tool to early predict patients at higher risk of death at ICU admission.

Predictive utilities of lipid traits, lipoprotein subfractions and other risk factors for incident diabetes: a machine learning approach in the Diabetes Prevention Program

INTRODUCTION

Although various lipid and non-lipid analytes measured by nuclear magnetic resonance (NMR) spectroscopy have been associated with type 2 diabetes, a structured comparison of the ability of NMR-derived biomarkers and standard lipids to predict individual diabetes risk has not been undertaken in larger studies nor among individuals at high risk of diabetes.

RESEARCH DESIGN AND METHODS

Cumulative discriminative utilities of various groups of biomarkers including NMR lipoproteins, related non-lipid biomarkers, standard lipids, and demographic and glycemic traits were compared for short-term (3.2 years) and long-term (15 years) diabetes development in the Diabetes Prevention Program, a multiethnic, placebo-controlled, randomized controlled trial of individuals with pre-diabetes in the USA (N=2590). Logistic regression, Cox proportional hazards model and six different hyperparameter-tuned machine learning algorithms were compared. The Matthews Correlation Coefficient (MCC) was used as the primary measure of discriminative utility.

RESULTS

Models with baseline NMR analytes and their changes did not improve the discriminative utility of simpler models including standard lipids or demographic and glycemic traits. Across all algorithms, models with baseline 2-hour glucose performed the best (max MCC=0.36). Sophisticated machine learning algorithms performed similarly to logistic regression in this study.

CONCLUSIONS

NMR lipoproteins and related non-lipid biomarkers were associated but did not augment discrimination of diabetes risk beyond traditional diabetes risk factors except for 2-hour glucose. Machine learning algorithms provided no meaningful improvement for discrimination compared with logistic regression, which suggests a lack of influential latent interactions among the analytes assessed in this study.

TRIAL REGISTRATION NUMBER

Diabetes Prevention Program: NCT00004992; Diabetes Prevention Program Outcomes Study: NCT00038727.

Artificial Intelligence and Early Detection of Pancreatic Cancer: 2020 Summative Review

ABSTRACT

Despite considerable research efforts, pancreatic cancer is associated with a dire prognosis and a 5-year survival rate of only 10%. Early symptoms of the disease are mostly nonspecific. The premise of improved survival through early detection is that more individuals will benefit from potentially curative treatment. Artificial intelligence (AI) methodology has emerged as a successful tool for risk stratification and identification in general health care. In response to the maturity of AI, Kenner Family Research Fund conducted the 2020 AI and Early Detection of Pancreatic Cancer Virtual Summit (www.pdac-virtualsummit.org) in conjunction with the American Pancreatic Association, with a focus on the potential of AI to advance early detection efforts in this disease. This comprehensive presummit article was prepared based on information provided by each of the interdisciplinary participants on one of the 5 following topics: Progress, Problems, and Prospects for Early Detection; AI and Machine Learning; AI and Pancreatic Cancer-Current Efforts; Collaborative Opportunities; and Moving Forward-Reflections from Government, Industry, and Advocacy. The outcome from the robust Summit conversations, to be presented in a future white paper, indicate that significant progress must be the result of strategic collaboration among investigators and institutions from multidisciplinary backgrounds, supported by committed funders.

SafeNET: Initial development and validation of a real-time tool for predicting mortality risk at the time of hospital transfer to a higher level of care

Background

Processes for transferring patients to higher acuity facilities lack a standardized approach to prognostication, increasing the risk for low value care that imposes significant burdens on patients and their families with unclear benefits. We sought to develop a rapid and feasible tool for predicting mortality using variables readily available at the time of hospital transfer.

Methods and findings

All work was carried out at a single, large, multi-hospital integrated healthcare system. We used a retrospective cohort for model development consisting of patients aged 18 years or older transferred into the healthcare system from another hospital, hospice, skilled nursing or other healthcare facility with an admission priority of direct emergency admit. The cohort was randomly divided into training and test sets to develop first a 54-variable, and then a 14-variable gradient boosting model to predict the primary outcome of all cause in-hospital mortality. Secondary outcomes included 30-day and 90-day mortality and transition to comfort measures only or hospice care. For model validation, we used a prospective cohort consisting of all patients transferred to a single, tertiary care hospital from one of the 3 referring hospitals, excluding patients transferred for myocardial infarction or maternal labor and delivery. Prospective validation was performed by using a web-based tool to calculate the risk of mortality at the time of transfer. Observed outcomes were compared to predicted outcomes to assess model performance.

The development cohort included 20,985 patients with 1,937 (9.2%) in-hospital mortalities, 2,884 (13.7%) 30-day mortalities, and 3,899 (18.6%) 90-day mortalities. The 14-variable gradient boosting model effectively predicted in-hospital, 30-day and 90-day mortality (c = 0.903 [95% CI:0.891–0.916]), c = 0.877 [95% CI:0.864–0.890]), and c = 0.869 [95% CI:0.857–0.881], respectively). The tool was proven feasible and valid for bedside implementation in a prospective cohort of 679 sequentially transferred patients for whom the bedside nurse calculated a SafeNET score at the time of transfer, taking only 4–5 minutes per patient with discrimination consistent with the development sample for in-hospital, 30-day and 90-day mortality (c = 0.836 [95%CI: 0.751–0.921], 0.815 [95% CI: 0.730–0.900], and 0.794 [95% CI: 0.725–0.864], respectively).

Conclusions

The SafeNET algorithm is feasible and valid for real-time, bedside mortality risk prediction at the time of hospital transfer. Work is ongoing to build pathways triggered by this score that direct needed resources to the patients at greatest risk of poor outcomes.

Clinician checklist for assessing suitability of machine learning applications in healthcare

Machine learning algorithms are being used to screen and diagnose disease, prognosticate and predict therapeutic responses. Hundreds of new algorithms are being developed, but whether they improve clinical decision making and patient outcomes remains uncertain. If clinicians are to use algorithms, they need to be reassured that key issues relating to their validity, utility, feasibility, safety and ethical use have been addressed. We propose a checklist of 10 questions that clinicians can ask of those advocating for the use of a particular algorithm, but which do not expect clinicians, as non-experts, to demonstrate mastery over what can be highly complex statistical and computational concepts. The questions are: (1) What is the purpose and context of the algorithm? (2) How good were the data used to train the algorithm? (3) Were there sufficient data to train the algorithm? (4) How well does the algorithm perform? (5) Is the algorithm transferable to new clinical settings? (6) Are the outputs of the algorithm clinically intelligible? (7) How will this algorithm fit into and complement current workflows? (8) Has use of the algorithm been shown to improve patient care and outcomes? (9) Could the algorithm cause patient harm? and (10) Does use of the algorithm raise ethical, legal or social concerns? We provide examples where an algorithm may raise concerns and apply the checklist to a recent review of diagnostic imaging applications. This checklist aims to assist clinicians in assessing algorithm readiness for routine care and identify situations where further refinement and evaluation is required prior to large-scale use.

Ethical Issues in Consent for the Reuse of Data in Health Data Platforms

Data platforms represent a new paradigm for carrying out health research. In the platform model, datasets are pooled for remote access and analysis, so novel insights for developing better stratified and/or personalised medicine approaches can be derived from their integration. If the integration of diverse datasets enables development of more accurate risk indicators, prognostic factors, or better treatments and interventions, this obviates the need for the sharing and reuse of data; and a platform-based approach is an appropriate model for facilitating this. Platform-based approaches thus require new thinking about consent. Here we defend an approach to meeting this challenge within the data platform model, grounded in: the notion of 'reasonable expectations' for the reuse of data; Waldron's account of 'integrity' as a heuristic for managing disagreement about the ethical permissibility of the approach; and the element of the social contract that emphasises the importance of public engagement in embedding new norms of research consistent with changing technological realities. While a social contract approach may sound appealing, however, it is incoherent in the context at hand. We defend a way forward guided by that part of the social contract which requires public approval for the proposal and argue that we have moral reasons to endorse a wider presumption of data reuse. However, we show that the relationship in question is not recognisably contractual and that the social contract approach is therefore misleading in this context. We conclude stating four requirements on which the legitimacy of our proposal rests.

Optimal dose and duration of enteral erythromycin as a prokinetic: A surgical intensive care experience

BACKGROUND

Enteral feeding has various advantages over parenteral feeding in critically ill patients. Acutely ill patients are at risk of developing enteral feeding intolerance. Prokinetic medications improve gastrointestinal mobility and enteral feed migration and absorption. Among the available prokinetic agents, erythromycin is the most potent. Erythromycin is used in different dosages and durations with variable efficacy. Intravenous erythromycin has an early and high rate of tachyphylaxis; hence, enteral route is preferred. Recently, the combination of prokinetic medications has been increasingly used because they accelerate the prokinetic action and decrease the adverse effects.

AIM

This study aimed to determine the optimal effective prokinetic dose and duration of administering enteral erythromycin in combination with metoclopramide in critically ill patients.

PATIENTS AND METHODS

This study has a prospective observation design. After obtaining permission from the medical research center of the institution, all patients in the surgical and trauma intensive care unit having enteral feed intolerance and those who were already on metoclopramide for 24 hour (h) were enrolled in the study. Patients' demographic data, diagnosis, surgical intervention, disease severity scores, erythromycin dose, duration of administration, any adverse effects, factors affecting erythromycin response, and outcome were recorded. All patients received 125 mg syrup erythromycin twice daily through a nasogastric tube (NGT). The NGT was clamped for 2 h, and half amount of previous enteral feeds was resumed. If the patient did not tolerate the feeds, the erythromycin dose was increased every 24 h in the increment of 250, 500, and 1000 mg (Figure 1). Statistical significance was considered at P <  0.05. A total of 313 patients were enrolled in the study. Majority of the patients were male, and the mean age was 45 years.

RESULTS

Majority (48.2%) of the patients (96) with feed intolerance were post laparotomy. Ninety percent (284) of the patients responded to prokinetic erythromycin therapy, and 54% received lower dose (125 mg twice daily). In addition, 14% had diarrhea, and none of these patients tested positive for Clostridium difficile toxin or multidrug resistance bacteria. The mean duration of erythromycin therapy was 4.98 days. The most effective prokinetic dose of erythromycin was 125 mg twice daily (P = 0.001). Erythromycin was significantly effective in patients with multiple organ dysfunction and shock (P = 0.001). Patients with high disease severity index and multiple organ dysfunction had significantly higher mortality (p < 0.05). Patients not responding to erythromycin therapy also had a significant higher mortality (p = 0.001).

CONCLUSION

Post-laparotomy patients had high enteral feed intolerance. Enteral erythromycin in combination with metoclopramide was effective in low dose and was required for short duration. Patients who did not tolerate feeds despite increasing dose of erythromycin had higher mortality.

Real-world implementation of precision psychiatry: Transdiagnostic risk calculator for the automatic detection of individuals at-risk of psychosis

BACKGROUND

Risk estimation models integrated into Electronic Health Records (EHRs) can deliver innovative approaches in psychiatry, but clinicians' endorsement and their real-world usability are unknown. This study aimed to investigate the real-world feasibility of implementing an individualised, transdiagnostic risk calculator to automatically screen EHRs and detect individuals at-risk for psychosis.

METHODS

Feasibility implementation study encompassing an in-vitro phase (March 2018 to May 2018) and in-vivo phase (May 2018 to April 2019). The in-vitro phase addressed implementation barriers and embedded the risk calculator (predictors: age, gender, ethnicity, index cluster diagnosis, age*gender) into the local EHR. The in-vivo phase investigated the real-world feasibility of screening individuals accessing secondary mental healthcare at the South London and Maudsley NHS Trust. The primary outcome was adherence of clinicians to automatic EHR screening, defined by the proportion of clinicians who responded to alerts from the risk calculator, over those contacted.

RESULTS

In-vitro phase: implementation barriers were identified/overcome with clinician and service user engagement, and the calculator was successfully integrated into the local EHR through the CogStack platform. In-vivo phase: 3722 individuals were automatically screened and 115 were detected. Clinician adherence was 74% without outreach and 85% with outreach. One-third of clinicians responded to the first email (37.1%) or phone calls (33.7%). Among those detected, cumulative risk of developing psychosis was 12% at six-month follow-up.

CONCLUSION

This is the first implementation study suggesting that combining precision psychiatry and EHR methods to improve detection of individuals with emerging psychosis is feasible. Future psychiatric implementation research is urgently needed.

ARDD 2020: from aging mechanisms to interventions

Aging is emerging as a druggable target with growing interest from academia, industry and investors. New technologies such as artificial intelligence and advanced screening techniques, as well as a strong influence from the industry sector may lead to novel discoveries to treat age-related diseases. The present review summarizes presentations from the 7th Annual Aging Research and Drug Discovery (ARDD) meeting, held online on the 1st to 4th of September 2020. The meeting covered topics related to new methodologies to study aging, knowledge about basic mechanisms of longevity, latest interventional strategies to target the aging process as well as discussions about the impact of aging research on society and economy. More than 2000 participants and 65 speakers joined the meeting and we already look forward to an even larger meeting next year. Please mark your calendars for the 8th ARDD meeting that is scheduled for the 31st of August to 3rd of September, 2021, at Columbia University, USA.

Association is not prediction: A landscape of confused reporting in diabetes - A systematic review

AIMS

Appropriate analysis of big data is fundamental to precision medicine. While statistical analyses often uncover numerous associations, associations themselves do not convey predictive value. Confusion between association and prediction harms clinicians, scientists, and ultimately, the patients. We analyzed published papers in the field of diabetes that refer to "prediction" in their titles. We assessed whether these articles report metrics relevant to prediction.

METHODS

A systematic search was undertaken using NCBI PubMed. Articles with the terms "diabetes" and "prediction" were selected. All abstracts of original research articles, within the field of diabetes epidemiology, were searched for metrics pertaining to predictive statistics. Simulated data was generated to visually convey the differences between association and prediction.

RESULTS

The search-term yielded 2,182 results. After discarding non-relevant articles, 1,910 abstracts were evaluated. Of these, 39% (n = 745) reported metrics of predictive statistics, while 61% (n = 1,165) did not. The top reported metrics of prediction were ROC AUC, sensitivity and specificity. Using the simulated data, we demonstrated that biomarkers with large effect sizes and low P values can still offer poor discriminative utility.

CONCLUSIONS

We demonstrate a landscape of confused reporting within the field of diabetes epidemiology where the term "prediction" is often incorrectly used to refer to association statistics. We propose guidelines for future reporting, and two major routes forward in terms of main analytic procedures and research goals: the explanatory route, which contributes to precision medicine, and the prediction route which contributes to personalized medicine.