Skip to main content
Advertisement
Browse Subject Areas
?

Click through the PLOS taxonomy to find articles in your field.

For more information about PLOS Subject Areas, click here.

  • Loading metrics

Dendrogram of transparent feature importance machine learning statistics to classify associations for heart failure: A reanalysis of a retrospective cohort study of the Medical Information Mart for Intensive Care III (MIMIC-III) database

  • Alexander A. Huang ,

    Contributed equally to this work with: Alexander A. Huang, Samuel Y. Huang

    Roles Conceptualization, Supervision, Validation, Visualization, Writing – original draft, Writing – review & editing

    Affiliation Department of MD Education, Northwestern University Feinberg School of Medicine, Chicago, IL, United States of America

  • Samuel Y. Huang

    Contributed equally to this work with: Alexander A. Huang, Samuel Y. Huang

    Roles Conceptualization, Formal analysis, Investigation, Methodology, Resources, Software, Supervision, Validation, Visualization, Writing – original draft, Writing – review & editing

    Huangs8@vcu.edu

    Affiliation Department of Internal Medicine, Virginia Commonwealth University School of Medicine, Richmond, VA, United States of America

Abstract

Background

There is a continual push for developing accurate predictors for Intensive Care Unit (ICU) admitted heart failure (HF) patients and in-hospital mortality.

Objective

The study aimed to utilize transparent machine learning and create hierarchical clustering of key predictors based off of model importance statistics gain, cover, and frequency.

Methods

Inclusion criteria of complete patient information for in-hospital mortality in the ICU with HF from the MIMIC-III database were randomly divided into a training (n = 941, 80%) and test (n = 235, 20%). A grid search was set to find hyperparameters. Machine Learning with XGBoost were used to predict mortality followed by feature importance with Shapely Additive Explanations (SHAP) and hierarchical clustering of model metrics with a dendrogram and heat map.

Results

Of the 1,176 heart failure ICU patients that met inclusion criteria for the study, 558 (47.5%) were males. The mean age was 74.05 (SD = 12.85). XGBoost model had an area under the receiver operator curve of 0.662. The highest overall SHAP explanations were urine output, leukocytes, bicarbonate, and platelets. Average urine output was 1899.28 (SD = 1272.36) mL/day with the hospital mortality group having 1345.97 (SD = 1136.58) mL/day and the group without hospital mortality having 1986.91 (SD = 1271.16) mL/day. The average leukocyte count in the cohort was 10.72 (SD = 5.23) cells per microliter. For the hospital mortality group the leukocyte count was 13.47 (SD = 7.42) cells per microliter and for the group without hospital mortality the leukocyte count was 10.28 (SD = 4.66) cells per microliter. The average bicarbonate value was 26.91 (SD = 5.17) mEq/L. Amongst the group with hospital mortality the average bicarbonate value was 24.00 (SD = 5.42) mEq/L. Amongst the group without hospital mortality the average bicarbonate value was 27.37 (SD = 4.98) mEq/L. The average platelet value was 241.52 platelets per microliter. For the group with hospital mortality the average platelet value was 216.21 platelets per microliter. For the group without hospital mortality the average platelet value was 245.47 platelets per microliter. Cluster 1 of the dendrogram grouped the temperature, platelets, urine output, Saturation of partial pressure of Oxygen (SPO2), Leukocyte count, lymphocyte count, bicarbonate, anion gap, respiratory rate, PCO2, BMI, and age as most similar in having the highest aggregate gain, cover, and frequency metrics.

Conclusion

Machine Learning models that incorporate dendrograms and heat maps can offer additional summaries of model statistics in differentiating factors between in patient ICU mortality in heart failure patients.

Introduction

Heart failure is a condition that affects a growing number of people and is one of the leading causes of death and hospitalization [1,2]. Patients with heart failure may need to spend more time in the hospital because there aren’t many options for managing their condition [35]. This is especially true for those with acute heart failure in the intensive care unit (ICU), where multiple underlying conditions may make their stay longer [6,7]. The financial burden of acute heart failure can have a significant impact on patient quality of life and is significant [8,9]. For many heart failure patients, particularly those with advanced organ dysfunction or severe complications, ICUs are necessary to provide advanced, high-tech, life-saving care [10,11]. ICUs have a high-intensity staffing model with high nurse and physician-to-patient ratios [12,13]. In the USA, approximately 10%–51% of hospitalized heart failure patients are admitted to an ICU [14,15]. ICU-admitted patients have significantly higher adjusted in-hospital mortality rates compared to those admitted to hospital wards only [16,17]. The in-hospital mortality rate for ICU-treated patients has been reported as 10.6%, whereas the rate for all HF patients is 4.0% [18,19]. Therefore, accurately predicting prognosis and providing intensive treatment with closer follow-up may be of greater benefit to ICU-admitted heart failure patients [20,21]. Although several in-hospital mortality prediction models are available, they lack model transparency and feature importance. Moreover, limited data are available on prediction models for ICU-admitted heart failure patients.

The use of machine learning in medicine for developing highly precise predictive models is on the rise [2224]. To achieve this, a common approach involves utilizing the XGBoost algorithm, which is known for its high accuracy, along with the transparent Shapely Additive Explanations (SHAP) algorithm to determine crucial covariates and their predictive direction [25,26]. In our research, we expanded upon this approach by integrating dendrograms and heatmaps to visually summarize covariates based on their gain, cover, and frequency. In the context of explainable machine learning, dendrograms provide additional insights by showing the relationships between variables based on their similarity, allowing for easier identification of important factors and potential interactions.

Table 1 compares the use of both SHAP and dendrograms, but ultimately how the use of both can provide additional information. SHAP values help us understand the impact of each feature on predictions, while dendrograms enhance model transparency by visualizing patterns and relationships among variables. Their application can support feature selection, model understanding, and decision-making processes in various domains. Eqs 1 and 2 similarly makes the above point using mathematical formulas. Eq 1 shows the general formula for calculating SHAP value as follows: (1)

thumbnail
Table 1. Comparison of the application of SHAP values and dendrograms in machine learning models.

https://doi.org/10.1371/journal.pone.0288819.t001

Where g(x’) represents the prediction of the explanation model for the specific coalition vector. Φ0 represents the intercept or bias term. Φj is the shaply value or feature attribution. represents the presence or absence of the feature in the model. The formulas deal primarily with each feature.

Eq 2 shows the general formula for calculating distances for dendrogram hierarchical clustering is as follows: (2)

Where dmn represents the distances between features m and n. i is the layer number and μ is the mean of the feature being compared. N is the number of layers. Utilizing dendrograms allow for comparing the relationship between multiple features in their relationship with one another.

We utilize the MIMIC-III (Medical Information Mart for Intensive Care III) dataset, a large-scale electronic health record database, to deploy our machine learning model. By using SHAP, we can not only identify important covariates, but also visualize the direction in which the model is predicting their effects. By grouping important covariates by the commonly used metrics of Gain, Cover, and Frequency, we can assess the performance of each covariate in the model.

Methods

The MIMIC-III database (V.1.4, 2016) contains de-identified data on 46,520 patients and 58,976 admissions to the ICU of the Beth Israel Deaconess Medical Center, Boston, USA, between June 1, 2001, and October 31, 2012. This publicly available critical care database provides comprehensive information on demographics, diagnoses, laboratory tests, medications, procedures, fluid balance, vital signs, radiology reports, and survival data. Approval to extract data from MIMIC-III after completing the National Institutes of Health Protecting Human Research Participants web-based training course was given via Certification Number: 28860101. The publicly available dataset was uploaded to Zenodo.

Dataset and cohort selection

Our investigation utilized data from the MIMIC-III (V.1.4, 2016), which was created to provide researchers with a comprehensive and freely accessible dataset of critical care patients to advance clinical research, patient care, and medical education. The database contains de-identified information on tens of thousands of ICU patients, including demographics, diagnoses, laboratory results, medications, procedures, and more. This allows researchers to study and analyze clinical outcomes, treatment patterns, and other critical care-related topics using real-world data.

For this study, we included adult patients (≥15 years old) diagnosed with heart failure (HF) identified through manual review of ICD-9 codes. Two researchers conducted the code review, and they excluded patients without an ICU record or missing left ventricular ejection fraction (LVEF) or N-terminal pro-brain natriuretic peptide (NT-proBNP) data. Data was read into R programming and any individual without an outcome for heart failure was excluded (N = 1,176).

Dependent variable

The study’s primary outcome was in-hospital mortality, which is defined as the survivors’ and non-survivors’ vital status at hospital discharge.

Independent variable

For this study, data were extracted from the following tables in the MIMIC-III dataset: ADMISSIONS, PATIENTS, ICUSTAYS, D_ICD DIAGNOSIS, DIAGNOSIS_ICD, LABEVENTS, D_LABIEVENTS, CHARTEVENTS, D_ITEMS, NOTEEVENTS, and OUTPUTEVENTS. The variables were selected based on their clinical relevance, general availability at the time of presentation, and previous studies.

The extracted data included demographic characteristics such as age, sex, ethnicity, weight, and height, as well as vital signs such as heart rate, systolic blood pressure, diastolic blood pressure, mean blood pressure, respiratory rate, body temperature, saturation pulse oxygen, and urine output in the first 24 hours. Comorbidities including hypertension, atrial fibrillation, ischemic heart disease, diabetes mellitus, depression, hypoferric anemia, hyperlipidemia, chronic kidney disease (CKD), and chronic obstructive pulmonary disease (COPD) were also recorded. Laboratory variables such as hematocrit, red blood cells, mean corpuscular hemoglobin, mean corpuscular hemoglobin concentration, mean corpuscular volume, red blood cell distribution width, platelet count, white blood cells, neutrophils, basophils, lymphocytes, prothrombin time, international normalized ratio, NT-proBNP, creatine kinase, creatinine, blood urea nitrogen, glucose, potassium, sodium, calcium, chloride, magnesium, the anion gap, bicarbonate, lactate, hydrogen ion concentration, partial pressure of CO2 in arterial blood, and left ventricular ejection fraction (LVEF) were also extracted.

The calculated mean value of variable data with multiple measurements collected throughout the hospital stay was used in the analysis. Variables with missing data are common in the MIMIC-III, however, eliminating patients with incomplete data can bias the study. Therefore, imputation is an important step in data preprocessing. All screening variables contained <25% missing values. Multiple imputation was done to handle missing values.

Model construction and statistical analysis

In univariate logistic models, the outcome of in-hospital mortality was used to identify covariates associated with each type. The machine learning model XGBoost was used because of its widespread use in the literature and improved predictive accuracy for healthcare predictions. Other studies using the NHANES cohort found that XGBoost offered the best balance between training efficiency, model accuracy, and transparency. The final set of model fit parameters (80:20) was calculated using a test and training set method. To determine the model’s fit, the area under the receiver operator characteristic curve (AUROC) was calculated.

Model feature importance statistics and SHAP visualization

The frequency, gain, and coverage were calculated for model covariates to identify risk factors associated with in-hospital mortality and they were ranked according to their gain. The feature’s relative contribution to the model’s predictions is shown by the Gain metric, while the feature’s total number of observations is shown by the Cover metric. On the other hand, frequency indicates how frequently a feature appears in the machine-learning model’s trees. Gain was chosen as the primary metric for ranking covariates because it is easy to understand and simple. Gain is the proportion of a given covariate’s influence on the final prediction. The strongest connections between the risk of hospital mortality and continuous covariates were visualized using SHAP explanations.

Dendrogram and heatmap creation based on gain, cover and frequency

Based on Gain, Cover, and Frequency, a dendrogram and heatmap were created. In the beginning, model covariates were ranked according to Gain, Cover, and Frequency in order to identify factors associated with hospital mortality. For each covariate, the Gain, Cover, and Frequency were calculated and sorted by value. The arranged covariate information was utilized to make a dendrogram that portrays the connection between different covariates in view of their comparability. Using Ward’s minimum variance method, the covariates were clustered based on their similarity to Gain, Cover, and Frequency for the dendrogram. The elbow method was used to figure out how many clusters were there, and k = 8 was chosen as the best number. The relationship between the covariates was then displayed on a heatmap alongside the hierarchical cluster. The goal of hierarchical cluster analysis is to create a tree diagram in which the items with the greatest degree of similarity are grouped together.

Results

Table 2 shows the 1,176-heart failure and ICU patients that met the inclusion criteria in this study. Of those, 159 (13.5%) individuals had in hospital mortality and 1,017 (86.5%) did not. There was 558 (47.4%) males in the total cohort with 80 (50.3%) in the hospital mortality cohort and 478 (47%) in the no hospital mortality cohort. Average age in the cohort that was 74.05 (SD = 12.85) with the hospital mortality group having an average age of 76.24 (SD = 13.22) and the group without hospital mortality having an average age of 73.71 (SD = 13.46). Average urine output was 1899.28 (SD = 1272.36) mL/day with the hospital mortality group having 1345.97 (SD = 1136.58) mL/day and the group without hospital mortality having 1986.91 (SD = 1271.16) mL/day. The average leukocyte count in the cohort was 10.72 (SD = 5.23) cells per microliter. For the hospital mortality group the leukocyte count was 13.47 (SD = 7.42) cells per microliter and for the group without hospital mortality the leukocyte count was 10.28 (SD = 4.66) cells per microliter. The average bicarbonate value was 26.91 (SD = 5.17) mEq/L. Amongst the group with hospital mortality the average bicarbonate value was 24.00 (SD = 5.42) mEq/L. Amongst the group without hospital mortality the average bicarbonate value was 27.37 (SD = 4.98) mEq/L. The average platelet value was 241.52 platelets per microliter. For the group with hospital mortality the average platelet value was 216.21 platelets per microliter. For the group without hospital mortality the average platelet value was 245.47 platelets per microliter.

Fig 1 displays the overall SHAP explanations for all the model covariates fit to predict in-hospital mortality to patients in heart failure patients in the intensive care unit. The model had a AUROC = 0.662.

thumbnail
Fig 1. Overall SHAP explanations.

SHAP explanations, purple color representing higher values of the covariate while yellow representing lower values of the covariate. X-axis is the change in log-odds for inpatient hospital mortality.

https://doi.org/10.1371/journal.pone.0288819.g001

Fig 2 depicts the SHAP visualizations for the four most significant continuous covariates based on overall SHAP explanations. Our findings indicate that increasing urine output up to 1,250 mL/day was associated with decreased inpatient hospital mortality in heart failure patients in the intensive care unit. Additionally, increased leukocyte count up to 20 cells per microliter was associated with increased inpatient hospital mortality in heart failure patients in the intensive care unit. Increased bicarbonate count up to 25 mEq/L was associated with decreased inpatient hospital mortality in heart failure patients in the intensive care unit. Increasing platelet count up to 200 platelets per microliter was associated with decreased inpatient hospital mortality in heart failure patients in the intensive care unit.

thumbnail
Fig 2. SHAP explanations for the Top 4 continuous covariates sorted by overall SHAP explanations.

SHAP explanations, covariate value on the x-axis, change in log-odds on the y-axis, red line represents the relationship between the covariate and log-odds for Hospital Mortality, each black dot represents an observation. Covariates: top left–Urine Output, top right–Leukocyte Count bottom left–Bicarbonate, bottom right–Platelets.

https://doi.org/10.1371/journal.pone.0288819.g002

Furthermore, Table 3 highlights the four top-ranked features by gain, which is a measure of the percentage contribution of the covariate to the overall model prediction. The most significant features were Bicarbonate (Gain = 6.7%), Platelets (Gain = 5.2%), Urine output (Gain = 5.1%), and temperature (Gain = 5%).

Fig 3 shows that in cluster 1 of the heatmap and dendrogram Temperature, platelets, urine output, Saturation of partial pressure of Oxygen (SPO2), Leukocyte count, lymphocyte count, bicarbonate, anion gap, respiratory rate, PCO2, BMI, and age were most similar in having high aggregate gain, cover, and frequency metrics.

thumbnail
Fig 3. Dendrogram and heatmap of covariates grouped by cover, frequency, and gain.

The colors of the dendrogram represent different groups and clusters of covariates based on their similarity in terms of the machine learning model metrics of cover, frequency, and gain. The length of the branches represents the degree of similarity between the covariates. Longer branches indicated greater differences between the covariates, and the closer together the branches are, the more similar covariates are. Colors of the heatmap show values of gain, cover, and frequency with red representing larger values and blue representing smaller values.

https://doi.org/10.1371/journal.pone.0288819.g003

Discussion

In this retrospective, cross sectional cohort of heart failure patients in the ICU, a machine learning model to mortality had a AUROC of 0.662. 1,176 patients with heart failure and ICU admission who met the study’s inclusion criteria. Out of these patients, 159 (13.5%) had in-hospital mortality, while 1,017 (86.5%) did not. Of the total cohort, 558 (47.4%) were males, with 80 (50.3%) in the hospital mortality group and 478 (47%) in the no hospital mortality group. The average age of the cohort was 74.05 (SD = 12.85), with the hospital mortality group having an average age of 76.24 (SD = 13.22) and the group without hospital mortality having an average age of 73.71 (SD = 13.46).

The model achieved an area under the receiver operating characteristic curve (AUROC) of 0.662, which suggests a moderate predictive accuracy. The study included 1,176 heart failure patients admitted to the ICU, out of which 13.5% experienced in-hospital mortality. The results indicate that age, urine output, leukocyte count, bicarbonate value, and platelet count were different between patients who experienced in-hospital mortality and those who did not and are consistent with other studies [21,2730]. Likewise, the findings that the hospital mortality group had a higher average age, lower urine output, higher leukocyte count, lower bicarbonate value, and lower platelet count compared to the group without hospital mortality are consistent with other studies [3134].

Machine learning models have been successful in heart failure and at detecting mortality. Heart failure mortality has been looked at in patients requiring different levels of care from step-down care, progressive care, and intensive care to different stages and with many different covariates [3538]. Researchers have utilized methods from logistic regression to machine learning [3944]. Within machine learning researchers are starting to utilize transparent methods for visualization [24,45].

A popular method that increases understandability of machine learning models is SHAP [22,46]. We use transparent machine learning methods to detect real signals that are in line with our current understandings as described in literature and clinical practice. The SHAP visualizations further support the increased predictive power of these non-parametric methods by demonstrating their ability to accurately capture the non-linear interactions between covariates, without overfitting the model to achieve greater accuracy.

We further introduce another way to visualize machine learning statistics. Dendrograms and heat maps are commonly used in various fields such as biology, ecology, genetics, and data science to visualize relationships and clusters [47]. Heat maps are particularly useful in condensing large amounts of information into a concise visual representation, and have been applied to gene expression, sequencing, geographic data, and population densities [48]. To better understand the complex relationships described by machine learning models, we proposed utilizing dendrograms and heat maps to describe the gain, cover, and frequency of covariates [49,50]. We found that the cluster produced by the dendrogram is similar to that produced by the SHAP value. Dendrograms provide additional insights in the context of explainable machine learning by visually displaying the relationship between variables based on their similarity, which can help identify important variables and uncover patterns in the data. In hierarchical clustering, variables are grouped together based on their similarity, with variables that are more similar being grouped together in the same cluster. Dendrograms can help identify clusters of variables that are highly correlated with each other, which can be useful in identifying variables that may be driving the model’s predictions or outcomes. Additionally, dendrograms can help identify variables that are not strongly correlated with any other variables, which can suggest that these variables may not be very informative for the model. Overall, dendrograms can provide a useful visual aid for interpreting the relationships between variables in an explainable machine learning context. The choice of using gain, cover, and frequency to cluster the variables was based on their importance in the XGBoost algorithm. These metrics provide valuable information about the predictive power of each variable and their contribution to the model’s overall performance. While two features with disparate tree positions may share similar gain scores, the clustering is based on overall similarity across all three metrics. Additionally, the dendrogram provides a visual representation of how the variables are clustered, which can aid in interpretation and further analysis. We acknowledge that there may be alternative approaches to clustering and welcome further discussion on this topic.

One direction of research with dendrograms is focused on exploring associations between outcomes and multiple variables, rather than just examining the relationship with a single variable. Dendrograms can be instrumental in identifying patterns and relationships among various factors, aiding in the creation of ordered sets or guidelines in fields such as medicine. For instance, in the context of heart failure, a dendrogram analysis can incorporate multiple features like RBC count, Troponin, BNP, Cr, BUN, and an echocardiogram, revealing the interconnections and similarities between these variables to inform the development of comprehensive and structured order sets for managing heart failure patients.

Depending on the specific problem, different statistics such as gain, cover, or frequency may hold more significance. Gain assesses the relative impact of each covariate on the model’s accuracy, which is especially relevant when high accuracy is paramount. Frequency measures the frequency with which a covariate appears in the model’s decision trees, providing insight into the trees’ patterns. Cover measures the number of instances of a given covariate, promoting generalizability. Since all three statistics hold importance and their relevance may vary depending on the problem, it is essential to develop a workflow that allows the modeler to evaluate how the machine learning algorithm weighs each metric. This facilitates feature selection.

Limitations

This machine-learning analysis has a retrospective nature, which may introduce bias. However, the potential bias was minimized by using training and testing sets to avoid overfitting. It is important to acknowledge this limitation. The use of SHAP visualizations can assist researchers in distinguishing whether the effects of each covariate are the result of true signal or noise, thereby reducing the risk of type-1 errors. Despite these limitations, we believe that machine-learning can serve as a useful preliminary measure in identifying potential risk factors. Subsequently, clinicians can evaluate these factors further based on the patient’s individual clinical presentation.

Conclusion

Machine learning models can find significant predictors for inpatient mortality in critically hospitalized heart failure patients. Feature importance with SHAP generates associations consistent with literature. Dendrograms and heat maps provide useful tools for model understandability.

References

  1. 1. Chang PP, Wruck LM, Shahar E, Rossi JS, Loehr LR, Russell SD, et al. Trends in Hospitalizations and Survival of Acute Decompensated Heart Failure in Four US Communities (2005–2014): ARIC Study Community Surveillance. Circulation. 2018;138(1):12–24. Epub 20180308. pmid:29519849; PubMed Central PMCID: PMC6030442.
  2. 2. Albright CM, Steiner J, Sienas L, Delgado C, Buber J. Main operating room deliveries for patients with high-risk cardiovascular disease. Open Heart. 2023;10(1). pmid:36787936; PubMed Central PMCID: PMC9930549.
  3. 3. Ponikowski P, Voors AA, Anker SD, Bueno H, Cleland JGF, Coats AJS, et al. 2016 ESC Guidelines for the diagnosis and treatment of acute and chronic heart failure: The Task Force for the diagnosis and treatment of acute and chronic heart failure of the European Society of Cardiology (ESC)Developed with the special contribution of the Heart Failure Association (HFA) of the ESC. Eur Heart J. 2016;37(27):2129–200. Epub 20160520. pmid:27206819.
  4. 4. Safavi KC, Dharmarajan K, Kim N, Strait KM, Li SX, Chen SI, et al. Variation exists in rates of admission to intensive care units for heart failure patients across hospitals in the United States. Circulation. 2013;127(8):923–9. Epub 20130125. pmid:23355624; PubMed Central PMCID: PMC3688061.
  5. 5. Belfiore A, Maurich A, Honjo O, Mazwi M, Jean-St-Michel E, Deng M, et al. Pitfalls and Possibilities of Ventricular Assist Device Support in Congenitally Corrected Transposition of the Great Arteries in Children. ASAIO J. 2023. Epub 20230307. pmid:36881646.
  6. 6. van Diepen S, Bakal JA, Lin M, Kaul P, McAlister FA, Ezekowitz JA. Variation in critical care unit admission rates and outcomes for patients with acute coronary syndromes or heart failure among high- and low-volume cardiac hospitals. J Am Heart Assoc. 2015;4(3):e001708. Epub 20150227. pmid:25725089; PubMed Central PMCID: PMC4392446.
  7. 7. Buscemi S, Davoli C, Trecarichi EM, Morrone HL, Tassone B, Buscemi C, et al. The three facets of the SARS-CoV-2 pandemic during the first two waves in the northern, central, and southern Italy. J Infect Public Health. 2023;16(4):520–5. Epub 20230207. pmid:36801631; PubMed Central PMCID: PMC9902343.
  8. 8. Wunsch H, Angus DC, Harrison DA, Collange O, Fowler R, Hoste EA, et al. Variation in critical care services across North America and Western Europe. Crit Care Med. 2008;36(10):2787–93,e1–9. pmid:18766102.
  9. 9. Cattin L, Ferrari F, Mongodi S, Pariani E, Bettini G, Daverio F, et al. Airways management in SARS-COV-2 acute respiratory failure: A prospective observational multi-center study. Med Intensiva. 2023;47(3):131–9. Epub 20230224. pmid:36855737; PubMed Central PMCID: PMC9950782.
  10. 10. Adams KF Jr, Fonarow GC, Emerman CL, LeJemtel TH, Costanzo MR, Abraham WT, et al. Characteristics and outcomes of patients hospitalized for heart failure in the United States: rationale, design, and preliminary observations from the first 100,000 cases in the Acute Decompensated Heart Failure National Registry (ADHERE). Am Heart J. 2005;149(2):209–16. pmid:15846257.
  11. 11. Edmiston EA, Hardin HK, Dolansky MA. Sleep Quality in the Advanced Heart Failure ICU. Clin Nurs Res. 2023:10547738231159045. Epub 20230306. pmid:36876721.
  12. 12. Fonarow GC, Adams KF Jr., Abraham WT, Yancy CW, Boscardin WJ, Adhere Scientific Advisory Committee SG, et al. Risk stratification for in-hospital mortality in acutely decompensated heart failure: classification and regression tree analysis. JAMA. 2005;293(5):572–80. pmid:15687312.
  13. 13. Gong F, Ai Y, Zhang L, Peng Q, Zhou Q, Gui C. Relationship between PaO(2)/FiO(2) and delirium in intensive care: A cross-sectional study. J Intensive Med. 2023;3(1):73–8. Epub 20221001. pmid:36789362; PubMed Central PMCID: PMC9923991.
  14. 14. Abraham WT, Fonarow GC, Albert NM, Stough WG, Gheorghiade M, Greenberg BH, et al. Predictors of in-hospital mortality in patients hospitalized for heart failure: insights from the Organized Program to Initiate Lifesaving Treatment in Hospitalized Patients with Heart Failure (OPTIMIZE-HF). J Am Coll Cardiol. 2008;52(5):347–56. pmid:18652942.
  15. 15. Grupper A, Chernomordik F, Herscovici R, Mazin I, Segev A, Beigel R, et al. The burden of heart failure in cardiac intensive care unit: a prospective 7 years analysis. ESC Heart Fail. 2023. Epub 20230217. pmid:36802123.
  16. 16. Peterson PN, Rumsfeld JS, Liang L, Albert NM, Hernandez AF, Peterson ED, et al. A validated risk score for in-hospital mortality in patients with heart failure from the American Heart Association get with the guidelines program. Circ Cardiovasc Qual Outcomes. 2010;3(1):25–32. Epub 20091208. pmid:20123668.
  17. 17. Guinot PG, Andrei S, Durand B, Martin A, Duclos V, Spitz A, et al. Balanced Nonopioid General Anesthesia With Lidocaine Is Associated With Lower Postoperative Complications Compared With Balanced Opioid General Anesthesia With Sufentanil for Cardiac Surgery With Cardiopulmonary Bypass: A Propensity Matched Cohort Study. Anesth Analg. 2023. Epub 20230210. pmid:36763521.
  18. 18. Yagyu T, Kumada M, Nakagawa T. Novel risk stratification with time course assessment of in-hospital mortality in patients with acute heart failure. PLoS One. 2017;12(11):e0187410. Epub 20171102. pmid:29095900; PubMed Central PMCID: PMC5667756.
  19. 19. Jentzer JC, Redfield MM, Killian J, Katz JN, Roger VL, Dunlay SM. Advanced Heart Failure in the Cardiac Intensive Care Unit: A Community-Based Study. JACC Heart Fail. 2023;11(2):252–4. pmid:36754533.
  20. 20. Ambale-Venkatesh B, Yang X, Wu CO, Liu K, Hundley WG, McClelland R, et al. Cardiovascular Event Prediction by Machine Learning: The Multi-Ethnic Study of Atherosclerosis. Circ Res. 2017;121(9):1092–101. Epub 20170809. pmid:28794054; PubMed Central PMCID: PMC5640485.
  21. 21. Keller K, Farmakis IT, Valerio L, Koelmel S, Wild J, Barco S, et al. Predisposing factors for admission to intensive care units of patients with COVID-19 infection-Results of the German nationwide inpatient sample. Front Public Health. 2023;11:1113793. Epub 20230215. pmid:36875366; PubMed Central PMCID: PMC9975593.
  22. 22. Hu X, Yang Z, Ma Y, Wang M, Liu W, Qu G, et al. Development and validation of a machine learning-based predictive model for secondary post-tonsillectomy hemorrhage. Front Surg. 2023;10:1114922. Epub 20230207. pmid:36824494; PubMed Central PMCID: PMC9941337.
  23. 23. Hu Y, Wu C, Meadows ME, Feng M. Pixel level spatial variability modeling using SHAP reveals the relative importance of factors influencing LST. Environ Monit Assess. 2023;195(3):407. Epub 20230216. pmid:36795252.
  24. 24. Huang AA, Huang SY. Increasing transparency in machine learning through bootstrap simulation and shapely additive explanations. PLoS One. 2023;18(2):e0281922. Epub 20230223. pmid:36821544; PubMed Central PMCID: PMC9949629.
  25. 25. Jiang J, Lu A, Ma X, Ouyang D, Williams RO 3rd. The applications of machine learning to predict the forming of chemically stable amorphous solid dispersions prepared by hot-melt extrusion. Int J Pharm X. 2023;5:100164. Epub 20230123. pmid:36798832; PubMed Central PMCID: PMC9925947.
  26. 26. Pan X, Feng T, Liu C, Savjani RR, Chin RK, Sharon Qi X. A survival prediction model via interpretable machine learning for patients with oropharyngeal cancer following radiotherapy. J Cancer Res Clin Oncol. 2023. Epub 20230218. pmid:36807760.
  27. 27. Kowsar R, Rahimi AM, Sroka M, Mansouri A, Sadeghi K, Bonakdar E, et al. Risk of mortality in COVID-19 patients: a meta- and network analysis. Sci Rep. 2023;13(1):2138. Epub 20230206. pmid:36747045; PubMed Central PMCID: PMC9901837.
  28. 28. Kulpins D, Pickney C, Garb M, Dickson TF, Young D, Patrinos ME, et al. Neonatal Intensive Care Unit Mixed Lipid Emulsion Use Associated With Reduced Cholestasis at Discharge in Surgical Patients. J Surg Res. 2023;287:1–7. Epub 20230222. pmid:36827839.
  29. 29. Belkin MN, Cifu AS, Pinney S. Management of Heart Failure. JAMA. 2022;328(13):1346–7. pmid:36107415.
  30. 30. Drazner MH. Left Ventricular Assist Devices in Advanced Heart Failure. JAMA. 2022;328(12):1207–9. pmid:36166048.
  31. 31. Kittleson MM. TRANSFORM-HF-Can We Close the Loop on Diuretics in Heart Failure? JAMA. 2023;329(3):211–3. pmid:36648482.
  32. 32. Mentz RJ, Anstrom KJ, Eisenstein EL, Sapp S, Greene SJ, Morgan S, et al. Effect of Torsemide vs Furosemide After Discharge on All-Cause Mortality in Patients Hospitalized With Heart Failure: The TRANSFORM-HF Randomized Clinical Trial. JAMA. 2023;329(3):214–23. pmid:36648467; PubMed Central PMCID: PMC9857435.
  33. 33. Reddy YNV, Koepp KE, Carter R, Win S, Jain CC, Olson TP, et al. Rate-Adaptive Atrial Pacing for Heart Failure With Preserved Ejection Fraction: The RAPID-HF Randomized Clinical Trial. JAMA. 2023. Epub 20230305. pmid:36871285; PubMed Central PMCID: PMC9986839.
  34. 34. Slomski A. Intensive Therapy Reduces Risks for Patients With Acute Heart Failure. JAMA. 2022;328(24):2387–8. pmid:36573977.
  35. 35. Linden A, Yarnold PR. Estimating causal effects for survival (time-to-event) outcomes by combining classification tree analysis and propensity score weighting. J Eval Clin Pract. 2018;24(2):380–7. Epub 20171212. pmid:29230910.
  36. 36. Neugebauer R, Schroeder EB, Reynolds K, Schmittdiel JA, Loes L, Dyer W, et al. Comparison of Mortality and Major Cardiovascular Events Among Adults With Type 2 Diabetes Using Human vs Analogue Insulins. JAMA Netw Open. 2020;3(1):e1918554. Epub 20200103. pmid:31977057; PubMed Central PMCID: PMC6991251.
  37. 37. Segar MW, Hall JL, Jhund PS, Powell-Wiley TM, Morris AA, Kao D, et al. Machine Learning-Based Models Incorporating Social Determinants of Health vs Traditional Models for Predicting In-Hospital Mortality in Patients With Heart Failure. JAMA Cardiol. 2022;7(8):844–54. pmid:35793094; PubMed Central PMCID: PMC9260645.
  38. 38. Segar MW, Jaeger BC, Patel KV, Nambi V, Ndumele CE, Correa A, et al. Development and Validation of Machine Learning-Based Race-Specific Models to Predict 10-Year Risk of Heart Failure: A Multicohort Analysis. Circulation. 2021;143(24):2370–83. Epub 20210413. pmid:33845593; PubMed Central PMCID: PMC9976274.
  39. 39. Mathen PG, Kumar KG, Mohan N, Sreekrishnan TP, Nair SB, Krishnan AK, et al. Prediction of Noninvasive Ventilation Failure in a Mixed Population Visiting the Emergency Department in a Tertiary Care Center in India. Indian J Crit Care Med. 2022;26(10):1115–9. pmid:36876205; PubMed Central PMCID: PMC9983674.
  40. 40. Sheng S, Xu FQ, Zhang YH, Huang Y. Charlson Comorbidity Index is correlated with all-cause readmission within six months in patients with heart failure: a retrospective cohort study in China. BMC Cardiovasc Disord. 2023;23(1):111. Epub 20230306. pmid:36879196; PubMed Central PMCID: PMC9987074.
  41. 41. Wu N, Li J, Xu X, Yuan Z, Yang L, Chen Y, et al. Prediction Model of New Onset Atrial Fibrillation in Patients with Acute Coronary Syndrome. Int J Clin Pract. 2023;2023:3473603. Epub 20230223. pmid:36874383; PubMed Central PMCID: PMC9981295.
  42. 42. Yao K, Wang J, Ma B, He L, Zhao T, Zou X, et al. A nomogram for predicting risk of death during hospitalization in elderly patients with Alzheimer’s disease at the time of admission. Front Neurol. 2023;14:1093154. Epub 20230216. pmid:36873432; PubMed Central PMCID: PMC9978216.
  43. 43. Yao X, Zhou J, Song L, Ren Y, Hu P, Liu D. A model-based meta analysis study of sodium glucose co-transporter-2 inhibitors. CPT Pharmacometrics Syst Pharmacol. 2023. Epub 20230308. pmid:36890732.
  44. 44. Yu YW, Huang Y, Zhao XM, Zhao L, Tian PC, Zhou Q, et al. The prognostic predictive value of the components of the PR interval in hospitalized patients with heart failure. BMC Cardiovasc Disord. 2023;23(1):119. Epub 20230308. pmid:36890463.
  45. 45. Zhou Y, Yao X, Han W, Wang Y, Li Z, Li Y. Distinguishing apathy and depression in older adults with mild cognitive impairment using text, audio, and video based on multiclass classification and shapely additive explanations. Int J Geriatr Psychiatry. 2022;37(11). pmid:36284449.
  46. 46. Hsu CT, Pai KC, Chen LC, Lin SH, Wu MJ. Machine Learning Models to Predict the Risk of Rapidly Progressive Kidney Disease and the Need for Nephrology Referral in Adult Patients with Type 2 Diabetes. Int J Environ Res Public Health. 2023;20(4). Epub 20230215. pmid:36834088; PubMed Central PMCID: PMC9967274.
  47. 47. Gamermann D, Montagud A, Conejero JA, Fernandez de Cordoba P, Urchueguia JF. Large scale evaluation of differences between network-based and pairwise sequence-alignment-based methods of dendrogram reconstruction. PLoS One. 2019;14(9):e0221631. Epub 20190905. pmid:31487289; PubMed Central PMCID: PMC6728023.
  48. 48. Nguyen TD, Schmidt B, Zheng Z, Kwoh CK. Efficient and Accurate OTU Clustering with GPU-Based Sequence Alignment and Dynamic Dendrogram Cutting. IEEE/ACM Trans Comput Biol Bioinform. 2015;12(5):1060–73. pmid:26451819.
  49. 49. Nthai D, Thibane VS, Gololo SS. Comparative Study of Abiotic Stress Factors on GC-MS-Detected Phytoconstituents of Aloe greatheadii var: davyana Using Heat Map and Hierarchical Clustering Dendrogram. Biochem Res Int. 2022;2022:5365024. Epub 20220105. pmid:35036008; PubMed Central PMCID: PMC8754611.
  50. 50. Ozturk B, Celik Y. Dendrogram for Anthropometric and Biomechanical Variables Causing Foot Deformities by Using Hierarchical Cluster Analysis: A Cross-Sectional Study. J Chiropr Med. 2022;21(2):108–15. Epub 20220404. pmid:35774634; PubMed Central PMCID: PMC9237588.