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

Key Concepts in Machine Learning and Clinical Applications in the Cardiac Intensive Care Unit

PURPOSE OF REVIEW

Artificial Intelligence (AI) technology will significantly alter critical care cardiology, from our understanding of diseases to the way in which we communicate with patients and colleagues. We summarize the potential applications of AI in the cardiac intensive care unit (CICU) by reviewing current evidence, future developments and possible challenges.

RECENT FINDINGS

Machine Learning (ML) methods have been leveraged to improve interpretation and discover novel uses for diagnostic tests such as the ECG and echocardiograms. ML-based dynamic risk stratification and prognostication may help optimize triaging and CICU discharge procedures. Latent class analysis and K-means clustering may reveal underlying disease sub-phenotypes within heterogeneous conditions such as cardiogenic shock and decompensated heart failure. AI technology may help enhance routine clinical care, facilitate medical education and training, and unlock individualized therapies for patients in the CICU. However, robust regulation and improved clinician understanding of AI is essential to overcome important practical and ethical challenges.

Development and evaluation of a model for predicting the risk of healthcare-associated infections in patients admitted to intensive care units

This retrospective study used 10 machine learning algorithms to predict the risk of healthcare-associated infections (HAIs) in patients admitted to intensive care units (ICUs). A total of 2,517 patients treated in the ICU of a tertiary hospital in China from January 2019 to December 2023 were included, of whom 455 (18.1%) developed an HAI. Data on 32 potential risk factors for infection were considered, of which 18 factors that were statistically significant on single-factor analysis were used to develop a machine learning prediction model using the synthetic minority oversampling technique (SMOTE). The main HAIs were respiratory tract infections (28.7%) and ventilator-associated pneumonia (25.0%), and were predominantly caused by gram-negative bacteria (78.8%). The CatBoost model showed good predictive performance (area under the curve: 0.944, and sensitivity 0.872). The 10 most important predictors of HAIs in this model were the Penetration Aspiration Scale score, Braden score, high total bilirubin level, female, high white blood cell count, Caprini Risk Score, Nutritional Risk Screening 2002 score, low eosinophil count, medium white blood cell count, and the Glasgow Coma Scale score. The CatBoost model accurately predicted the occurrence of HAIs and could be used in clinical practice.

Interpretability-based machine learning for predicting the risk of death from pulmonary inflammation in Chinese intensive care unit patients

Objective

The objective of this research was to create a machine learning predictive model that could be easily interpreted in order to precisely determine the risk of premature death in patients receiving intensive care after pulmonary inflammation.

Methods

In this study, information from the China intensive care units (ICU) Open Source database was used to examine data from 2790 patients who had infections between January 2019 and December 2020. A 7:3 ratio was used to randomly assign the whole patient population to training and validation groups. This study used six machine learning techniques: logistic regression, random forest, gradient boosting tree, extreme gradient boosting tree (XGBoost), multilayer perceptron, and K-nearest neighbor. A cross-validation grid search method was used to search the parameters in each model. Eight metrics were used to assess the models' performance: accuracy, precision, recall, F1 score, area under the curve (AUC) value, Brier score, Jordon's index, and calibration slope. The machine methods were ranked based on how well they performed in each of these metrics. The best-performing models were selected for interpretation using both the Shapley Additive exPlanations (SHAP) and Local interpretable model-agnostic explanations (LIME) interpretable techniques.

Results

A subset of the study cohort's patients (120/1668, or 7.19%) died in the hospital following screening for inclusion and exclusion criteria. Using a cross-validated grid search to evaluate the six machine learning techniques, XGBoost showed good discriminative ability, achieving an accuracy score of 0.889 (0.874-0.904), precision score of 0.871 (0.849-0.893), recall score of 0.913 (0.890-0.936), F1 score of 0.891 (0.876-0.906), and AUC of 0.956 (0.939-0.973). Additionally, XGBoost exhibited excellent performance with a Brier score of 0.050, Jordon index of 0.947, and calibration slope of 1.074. It was also possible to create an interactive internet page using the XGBoost model.

Conclusion

By identifying patients at higher risk of early mortality, machine learning-based mortality risk prediction models have the potential to significantly improve patient care by directing clinical decision making and enabling early detection of survival and mortality issues in patients with pulmonary inflammation disease.

Public and Research Interest in Telemedicine From 2017 to 2022: Infodemiology Study of Google Trends Data and Bibliometric Analysis of Scientific Literature

BACKGROUND

Telemedicine offers a multitude of potential advantages, such as enhanced health care accessibility, cost reduction, and improved patient outcomes. The significance of telemedicine has been underscored by the COVID-19 pandemic, as it plays a crucial role in maintaining uninterrupted care while minimizing the risk of viral exposure. However, the adoption and implementation of telemedicine have been relatively sluggish in certain areas. Assessing the level of interest in telemedicine can provide valuable insights into areas that require enhancement.

OBJECTIVE

The aim of this study is to provide a comprehensive analysis of the level of public and research interest in telemedicine from 2017 to 2022 and also consider any potential impact of the COVID-19 pandemic.

METHODS

Google Trends data were retrieved using the search topics "telemedicine" or "e-health" to assess public interest, geographic distribution, and trends through a joinpoint regression analysis. Bibliographic data from Scopus were used to chart publications referencing the terms "telemedicine" or "eHealth" (in the title, abstract, and keywords) in terms of scientific production, key countries, and prominent keywords, as well as collaboration and co-occurrence networks.

RESULTS

Worldwide, telemedicine generated higher mean public interest (relative search volume=26.3%) compared to eHealth (relative search volume=17.6%). Interest in telemedicine remained stable until January 2020, experienced a sudden surge (monthly percent change=95.7%) peaking in April 2020, followed by a decline (monthly percent change=-22.7%) until August 2020, and then returned to stability. A similar trend was noted in the public interest regarding eHealth. Chile, Australia, Canada, and the United States had the greatest public interest in telemedicine. In these countries, moderate to strong correlations were evident between Google Trends and COVID-19 data (ie, new cases, new deaths, and hospitalized patients). Examining 19,539 original medical articles in the Scopus database unveiled a substantial rise in telemedicine-related publications, showing a total increase of 201.5% from 2017 to 2022 and an average annual growth rate of 24.7%. The most significant surge occurred between 2019 and 2020. Notably, the majority of the publications originated from a single country, with 20.8% involving international coauthorships. As the most productive country, the United States led a cluster that included Canada and Australia as well. European, Asian, and Latin American countries made up the remaining 3 clusters. The co-occurrence network categorized prevalent keywords into 2 clusters, the first cluster primarily focused on applying eHealth, mobile health (mHealth), or digital health to noncommunicable or chronic diseases; the second cluster was centered around the application of telemedicine and telehealth within the context of the COVID-19 pandemic.

CONCLUSIONS

Our analysis of search and bibliographic data over time and across regions allows us to gauge the interest in this topic, offer evidence regarding potential applications, and pinpoint areas for additional research and awareness-raising initiatives.

Chinese experts' consensus on the application of intensive care big data

The development of intensive care medicine is inseparable from the diversified monitoring data. Intensive care medicine has been closely integrated with data since its birth. Critical care research requires an integrative approach that embraces the complexity of critical illness and the computational technology and algorithms that can make it possible. Considering the need of standardization of application of big data in intensive care, Intensive Care Medicine Branch of China Health Information and Health Care Big Data Society, Standard Committee has convened expert group, secretary group and the external audit expert group to formulate Chinese Experts' Consensus on the Application of Intensive Care Big Data (2022). This consensus makes 29 recommendations on the following five parts: Concept of intensive care big data, Important scientific issues, Standards and principles of database, Methodology in solving big data problems, Clinical application and safety consideration of intensive care big data. The consensus group believes this consensus is the starting step of application big data in the field of intensive care. More explorations and big data based retrospective research should be carried out in order to enhance safety and reliability of big data based models of critical care field.

Association between quality of governance, antibiotic consumption, and antimicrobial resistance: an analysis of Italian regions

BACKGROUND

Emerging research has provided evidence suggesting the potential influence of governance on the development and spread of antimicrobial resistance (AMR), accounting for significant disparities observed both between and within countries. In our study, we conducted an ecological analysis to investigate the relationship between governance quality, antibiotic consumption, and AMR across Italian regions.

METHODS

By leveraging data from three distinct sources at the regional level, we compiled a comprehensive dataset comprising: AMR proportions for three specific pathogen-antibiotic combinations in the year 2021, antibiotic consumption data for systemic use in the year 2020, and the 2021 European Quality of Government Index (EQI) and its corresponding pillars. Employing mediation analysis, we investigated the potential mediating role of antibiotic consumption in the association between the EQI and an average measure of AMR.

RESULTS

Our analysis revealed substantial variation in the percentages of AMR across different regions in Italy, demonstrating a discernible North-to-South gradient concerning both antibiotic usage and governance quality. The EQI exhibited a statistically significant negative correlation with both antibiotic consumption and AMR percentages, encompassing both specific combinations and their average value. Regions characterized by higher levels of governance quality consistently displayed lower values of antibiotic consumption and AMR, while regions with lower governance quality tended to exhibit higher levels of antibiotic use and AMR. Furthermore, we observed a significant total effect of the EQI on average AMR (β = - 0.97; CI - 1.51; - 0.43). Notably, this effect was found to be mediated by antibiotic consumption, as evidenced by a significant indirect effect (β = - 0.89; CI - 1.45; - 0.32).

CONCLUSIONS

These findings draw attention to the regional disparities observed in AMR levels, antibiotic consumption patterns, and governance quality in Italy. Our study also highlights the mediating role of antibiotic consumption in the relationship between governance quality and AMR. This underscores the significance of implementing focused interventions and policies aimed at improving governance quality and promoting responsible antibiotic use.

The intertwining of healthcare-associated infections and COVID-19 in Italian intensive care units: an analysis of the SPIN-UTI project from 2006 to 2021

BACKGROUND

Although healthcare-associated infections (HAIs) pose an extraordinary burden on public health, the impact of coronavirus disease 2019 (COVID-19) is still a matter of debate.

AIM

To describe trends of HAIs in Italian intensive care units (ICUs) from 2006 to 2021, and to compare characteristics and outcomes of patients with or without COVID-19.

METHODS

We evaluated patients participating in the 'Italian Nosocomial Infections Surveillance in Intensive Care Units' (SPIN-UTI) project, who were admitted to ICUs for more than 48 h. Data regarding diagnosis, clinical conditions, therapies, treatments and outcomes of COVID-19 patients were also collected.

FINDINGS

From a total of 21,523 patients from 2006 to 2021, 3485 (16.2%) presented at least one HAI. We observed an increasing trend for both the incidence of patients with HAI and the incidence density of HAIs (P-trend <0.001). Compared with the pre-pandemic period, the incidence density of HAIs increased by about 15% in 2020-2021, with pneumoniae being the greatest contributors to this increase (P-trend <0.001). Moreover, incidence of HAIs was higher in ICUs dedicated to COVID-19 patients (P<0.001), who showed a greater risk of HAIs and death than patients without COVID-19 (P-values <0.001). Accordingly, the mortality in ICUs increased over the years and doubled during the pandemic (P-trend <0.001). Notably, co-infected patients had higher mortality (75.2%) than those with COVID-19 (66.2%) or HAI (39.9%) alone, and those without any infection (23.2%).

CONCLUSIONS

Our analysis provides useful insight into whether and how the COVID-19 pandemic influenced HAI incidence and death in Italian ICUs, highlighting the need for evaluation of the long-term effects of the pandemic.

Factors Affecting the Mortality Rate in Non-COVID-19 Intensive Care Unit Patients During the COVID-19 Pandemic in Cyprus: A Retrospective Cohort Study

INTRODUCTION

Mortality statistics constitute a pivotal element in informing public health policymaking in critical care settings. Mortality rates exhibit temporal variability, and their quantification is susceptible to well-established biases that have been exacerbated in the backdrop of the COVID-19 pandemic. A multitude of factors contribute to the process of patients' outcomes within the intensive care unit (ICU) setting. The primary aim of this study is to compare the mortality rate observed during the initial and subsequent phases of the COVID-19 pandemic in non-COVID-19 patient cohorts. Secondary objectives encompass evaluating the demographic and clinical factors and admission times to the ICU as an independent predictor affecting mortality.

METHODS AND MATERIALS

 A retrospective investigation of the data gathered from 1127 non-COVID-19 patients admitted to an ICU situated in Nicosia, Cyprus between March 2020 and December 2022 was performed. We divided the study period into two distinct timeframes. The first period spanned from the onset of the COVID-19 pandemic up until January 2021, coinciding with the relaxation of COVID-related restrictions. The second period was defined as the period when restrictions were not applied. The time of admission to the ICU is categorized as either off-hours or business hours. We recorded various patient characteristics, including age, gender, Acute Physiology and Chronic Health Evaluation II (APACHE II), Glasgow Coma Scale (GCS), Sequential Organ Failure Assessment (SOFA) scores, hospitalization duration, discharge details, mortality events with precise timestamps and primary diagnosis for admission. Multivariate logistic regression analysis was performed with these characteristics to predict the likelihood of mortality.

RESULTS

 This study included 632 males (56.1%) and 495 females (43.9%). Within the patient cohort, 653 patients (57.9%) were discharged from the ICU, while 474 patients (42.1%) experienced mortality during their ICU stay. No significant correlation was found whether patients were admitted to ICU during the first or second period of the COVID-19 pandemic. There was a significant difference in the comparison of outcomes within the ICU between the off-hours and business hours (p=0.001). A total of 329 of 618 (53.2%) patients admitted in off-hours and 145 of 509 (28.4%) patients admitted in business hours died. Moreover, the mean GCS, APACHE II and SOFA scores were higher in patients admitted during off-hours. APACHE II score (OR: 1.11, 95% CI: 1.06 to 1.15, p<0.01), SOFA (OR: 1.21, 95% CI: 1.10 to 1.31, p<0.01) and GCS (OR: 0.88, 95% CI: 0.84 to 0.92, p<0.01) scores and admission to the ICU in off-hours 2.63 (1.91-3.67) were significantly associated with mortality.

CONCLUSION

 The results of this retrospective cohort analysis have shown that the mortality rate was higher in non-COVID-19 patients admitted to ICU during off-hours compared to those admitted during business hours. However, no significant difference was found in the mortality rate between the admissions during the first and second periods of the COVID-19 pandemic.

An observational study on factors associated with ICU mortality in Covid-19 patients and critical review of the literature

The novel pandemic caused by SARS-CoV-2 has been associated with increased burden on healthcare system. Recognizing the variables that independently predict death in COVID-19 is of great importance. The study was carried out prospectively in a single ICU in northern Greece. It was based on the collection of data during clinical practice in 375 adult patients who were tested positive for SARS-CoV-2 between April 2020 and February 2022. All patients were intubated due to acute respiratory insufficiency and received Invasive Mechanical Ventilation. The primary outcome was ICU mortality. Secondary outcomes were 28-day mortality and independent predictors of mortality at 28 days and during ICU hospitalization. For continuous variables with normal distribution, t-test was used for means comparison between two groups and one-way ANOVA for multiple comparisons. When the distribution was not normal, comparisons were performed using the Mann-Whitney test. Comparisons between discrete variables were made using the x² test, whereas the binary logistic regression was employed for the definition of factors affecting survival inside the ICU and after 28 days. Of the total number of patients intubated due to COVID-19 during the study period, 239 (63.7%) were male. Overall, the ICU survival was 49.6%, whereas the 28-day survival reached 46.9%. The survival rates inside the ICU for the four main viral variants were 54.9%, 50.3%, 39.7% and 50% for the Alpha, Beta, Delta and Omicron variants, respectively. Logistic regressions for outcome revealed that the following parameters were independently associated with ICU survival: wave, SOFA @day1, Remdesivir use, AKI, Sepsis, Enteral Insufficiency, Duration of ICU stay and WBC. Similarly, the parameters affecting the 28-days survival were: duration of stay in ICU, SOFA @day1, WBC, Wave, AKI and Enteral Insufficiency. In this observational cohort study of critically ill COVID-19 patients we report an association between mortality and the wave sequence, SOFA score on admission, the use of Remdesivir, presence of AKI, presence of gastrointestinal failure, sepsis and WBC levels. Strengths of this study are the large number of critically ill COVID-19 patients included, and the comparison of the adjusted mortality rates between pandemic waves within a two year-study period.

Machine Learning Identification of Obstructive Sleep Apnea Severity through the Patient Clinical Features: A Retrospective Study

OBJECTIVES

To evaluate the role of clinical scores assessing the risk of disease severity in patients with clinical suspicion of obstructive sleep apnea syndrome (OSA). The hypothesis was tested by applying artificial intelligence (AI) to demonstrate its effectiveness in distinguishing between mild-moderate OSA and severe OSA risk.

METHODS

A support vector machine model (SVM) was developed from the samples included in the analysis (N = 498), and they were split into 75% for training (N = 373) with the remaining for testing (N = 125). Two diagnostic thresholds were selected for OSA severity: mild to moderate (apnea-hypopnea index (AHI) ≥ 5 events/h and AHI < 30 events/h) and severe (AHI ≥ 30 events/h). The algorithms were trained and tested to predict OSA patient severity.

RESULTS

The sensitivity and specificity for the SVM model were 0.93 and 0.80 with an accuracy of 0.86; instead, the logistic regression full mode reported a value of 0.74 and 0.63, respectively, with an accuracy of 0.68. After backward stepwise elimination for features selection, the reduced logistic regression model demonstrated a sensitivity and specificity of 0.79 and 0.56, respectively, and an accuracy of 0.67.

CONCLUSION

Artificial intelligence could be applied to patients with symptoms related to OSA to identify individuals with a severe OSA risk with clinical-based algorithms in the OSA framework.

Applying an Improved Stacking Ensemble Model to Predict the Mortality of ICU Patients with Heart Failure

Cardiovascular diseases have been identified as one of the top three causes of death worldwide, with onset and deaths mostly due to heart failure (HF). In ICU, where patients with HF are at increased risk of death and consume significant medical resources, early and accurate prediction of the time of death for patients at high risk of death would enable them to receive appropriate and timely medical care. The data for this study were obtained from the MIMIC-III database, where we collected vital signs and tests for 6699 HF patient during the first 24 h of their first ICU admission. In order to predict the mortality of HF patients in ICUs more precisely, an integrated stacking model is proposed and applied in this paper. In the first stage of dataset classification, the datasets were subjected to first-level classifiers using RF, SVC, KNN, LGBM, Bagging, and Adaboost. Then, the fusion of these six classifier decisions was used to construct and optimize the stacked set of second-level classifiers. The results indicate that our model obtained an accuracy of 95.25% and AUROC of 82.55% in predicting the mortality rate of HF patients, which demonstrates the outstanding capability and efficiency of our method. In addition, the results of this study also revealed that platelets, glucose, and blood urea nitrogen were the clinical features that had the greatest impact on model prediction. The results of this analysis not only improve the understanding of patients' conditions by healthcare professionals but allow for a more optimal use of healthcare resources.

Diabetic Retinopathy Detection from Fundus Images of the Eye Using Hybrid Deep Learning Features

Diabetic Retinopathy (DR) is a medical condition present in patients suffering from long-term diabetes. If a diagnosis is not carried out at an early stage, it can lead to vision impairment. High blood sugar in diabetic patients is the main source of DR. This affects the blood vessels within the retina. Manual detection of DR is a difficult task since it can affect the retina, causing structural changes such as Microaneurysms (MAs), Exudates (EXs), Hemorrhages (HMs), and extra blood vessel growth. In this work, a hybrid technique for the detection and classification of Diabetic Retinopathy in fundus images of the eye is proposed. Transfer learning (TL) is used on pre-trained Convolutional Neural Network (CNN) models to extract features that are combined to generate a hybrid feature vector. This feature vector is passed on to various classifiers for binary and multiclass classification of fundus images. System performance is measured using various metrics and results are compared with recent approaches for DR detection. The proposed method provides significant performance improvement in DR detection for fundus images. For binary classification, the proposed modified method achieved the highest accuracy of 97.8% and 89.29% for multiclass classification.

Machine Learning and Antibiotic Management

Machine learning and cluster analysis applied to the clinical setting of an intensive care unit can be a valuable aid for clinical management, especially with the increasing complexity of clinical monitoring. Providing a method to measure clinical experience, a proxy for that automatic gestalt evaluation that an experienced clinician sometimes effortlessly, but often only after long, hard consideration and consultation with colleagues, relies upon for decision making, is what we wanted to achieve with the application of machine learning to antibiotic therapy and clinical monitoring in the present work. This is a single-center retrospective analysis proposing methods for evaluation of vitals and antimicrobial therapy in intensive care patients. For each patient included in the present study, duration of antibiotic therapy, consecutive days of treatment and type and combination of antimicrobial agents have been assessed and considered as single unique daily record for analysis. Each parameter, composing a record was normalized using a fuzzy logic approach and assigned to five descriptive categories (fuzzy domain sub-sets ranging from “very low” to “very high”). Clustering of these normalized therapy records was performed, and each patient/day was considered to be a pertaining cluster. The same methodology was used for hourly bed-side monitoring. Changes in patient conditions (monitoring) can lead to a shift of clusters. This can provide an additional tool for assessing progress of complex patients. We used Fuzzy logic normalization to descriptive categories of parameters as a form nearer to human language than raw numbers.

Predictive Analysis of Healthcare-Associated Blood Stream Infections in the Neonatal Intensive Care Unit Using Artificial Intelligence: A Single Center Study

Background: Neonatal infections represent one of the six main types of healthcare-associated infections and have resulted in increasing mortality rates in recent years due to preterm births or problems arising from childbirth. Although advances in obstetrics and technologies have minimized the number of deaths related to birth, different challenges have emerged in identifying the main factors affecting mortality and morbidity. Dataset characterization: We investigated healthcare-associated infections in a cohort of 1203 patients at the level III Neonatal Intensive Care Unit (ICU) of the “Federico II” University Hospital in Naples from 2016 to 2020 (60 months). Methods: The present paper used statistical analyses and logistic regression to identify an association between healthcare-associated blood stream infection (HABSIs) and the available risk factors in neonates and prevent their spread. We designed a supervised approach to predict whether a patient suffered from HABSI using seven different artificial intelligence models. Results: We analyzed a cohort of 1203 patients and found that birthweight and central line catheterization days were the most important predictors of suffering from HABSI. Conclusions: Our statistical analyses showed that birthweight and central line catheterization days were significant predictors of suffering from HABSI. Patients suffering from HABSI had lower gestational age and birthweight, which led to longer hospitalization and umbilical and central line catheterization days than non-HABSI neonates. The predictive analysis achieved the highest Area Under Curve (AUC), accuracy and F1-macro score in the prediction of HABSIs using Logistic Regression (LR) and Multi-layer Perceptron (MLP) models, which better resolved the imbalanced dataset (65 infected and 1038 healthy).

U-Shape Relationship between Plasma Leucine Level and Mortality in the Intensive Care Unit

Patients in the intensive care unit (ICU) are at high risk of mortality which is not well predicted. Previous studies noted that leucine has prognostic value in a variety of diseases. This study investigated whether leucine concentration was a useful biomarker of metabolic and nutritional status and 6-month mortality in ICU. We recruited 454 subjects admitted to ICU (348 and 106 in the initiation and validation cohorts, respectively) with an acute physiology and chronic health evaluation (APACHE II) score ≥ 15. We measured plasma leucine concentrations, traditional biomarkers, and calculated APACHE II and sequential organ failure assessment (SOFA) scores. Leucine levels were weakly correlated with albumin, prealbumin, and transferrin levels (r = 0.30, 0.12, and 0.15, p = 0.001, 0.029, and 0.007, respectively). During follow-up, 116 (33.3%) patients died. Compared to patients with leucine levels between 109 and 174 μM, patients with leucine > 174 μM or <109 μM had a lower cumulative survival rate. Death was also associated with age, higher APACHE II and SOFA scores, C-reactive protein, and longer stays in the ICU, but with lower albumin, prealbumin, and transferrin. Patients with leucine levels > 174 μM had higher alanine aminotransferase levels, but no significant differences in other variables; patients with leucine levels < 109 μM had higher APACHE II and SOFA scores, higher incidence of using inotropic agents, longer ICU and hospital stays, but lower albumin and transferrin levels. Multivariable analysis demonstrated that leucine > 174 μM was an independent predictor of mortality, especially early mortality. However, among patients who stayed in ICU longer than two weeks, leucine < 109 μM was an independent predictor of mortality. In addition, leucine < 109 μM was associated with worse ventilator weaning profiles. These findings were similar in the validation cohort. Our study demonstrated a U-shape relationship between leucine levels and mortality rate in ICU.

Disentangling the Association of Hydroxychloroquine Treatment with Mortality in Covid-19 Hospitalized Patients through Hierarchical Clustering

The efficacy of hydroxychloroquine (HCQ) in treating SARS-CoV-2 infection is harshly debated, with observational and experimental studies reporting contrasting results. To clarify the role of HCQ in Covid-19 patients, we carried out a retrospective observational study of 4,396 unselected patients hospitalized for Covid-19 in Italy (February–May 2020). Patients' characteristics were collected at entry, including age, sex, obesity, smoking status, blood parameters, history of diabetes, cancer, cardiovascular and chronic pulmonary diseases, and medications in use. These were used to identify subtypes of patients with similar characteristics through hierarchical clustering based on Gower distance. Using multivariable Cox regressions, these clusters were then tested for association with mortality and modification of effect by treatment with HCQ. We identified two clusters, one of 3,913 younger patients with lower circulating inflammation levels and better renal function, and one of 483 generally older and more comorbid subjects, more prevalently men and smokers. The latter group was at increased death risk adjusted by HCQ (HR[CI95%] = 3.80[3.08-4.67]), while HCQ showed an independent inverse association (0.51[0.43-0.61]), as well as a significant influence of cluster∗HCQ interaction (p < 0.001). This was driven by a differential association of HCQ with mortality between the high (0.89[0.65-1.22]) and the low risk cluster (0.46[0.39-0.54]). These effects survived adjustments for additional medications in use and were concordant with associations with disease severity and outcome. These findings suggest a particularly beneficial effect of HCQ within low risk Covid-19 patients and may contribute to clarifying the current controversy on HCQ efficacy in Covid-19 treatment.