Predictive divergence in machine learning models for clinical mortality risk: A multicohort study of covid-19 patients

Background Machine learning (ML) algorithms are increasingly used in healthcare to support clinical decision-making. While models with similar overall performance are often considered interchangeable for deployment, they may produce divergent predictions, a phenomenon known as algorithmic multiplicity. In such cases, the choice of algorithm may introduce bias. This study investigates the impacts of algorithmic multiplicity in mortality prediction and assesses the influence of patient characteristics on model decisions. Methods A cohort of 4,337 adult patients (≥18 years) with RT-PCR–confirmed covid-19 from five tertiary care hospitals in Brazil was followed from March to August 2020. Five popular ML models for structured data were trained on demographic and laboratory data collected at early hospital admission to predict in-hospital mortality. Model performance, feature importance, and algorithmic prediction similarity were evaluated. Feature distributions were compared between patients correctly or incorrectly classified by all models using paired t-tests or Mann–Whitney U tests, as applicable, at the 5% significance level. Subgroup performance differences were assessed using 10-fold cross-validation applied to five k-means–delineated clusters, compared by one-way ANOVA. Within-cluster predictive divergence was assessed within a 95% confidence interval. Results All models achieved high overall predictive performance (µ = 0.855, σ² = 0.0072). However, the comparison of individual-level predictions revealed substantial heterogeneity, with pairwise prediction correlations ranging from R² = 0.56 to 0.80. Unsupervised k-means clustering identified five clinically distinct patient subgroups with mortality rates ranging from 22% to 80%, within which model performance varied significantly (F = 73.18, p < 0.001). Notably, TabPFN and LightGBM showed superior performance in the “Anemia” cluster, whereas TabPFN underperformed in the “Immunodeficient” cluster (95% CI). Conclusions This study demonstrates that ML models with similar overall performance can yield substantially divergent predictions at both the individual and subgroup levels, and that no single algorithm consistently outperforms others across all patient subgroups. These findings highlight the limitations of relying solely on global performance metrics and underscore the need for context-aware evaluation of ML models in heterogeneous clinical populations.

Predicting negative self-rated oral health in adults using machine learning: A longitudinal study in Southern Brazil

Global performance of machine learning models to predict all-cause mortality: systematic review and meta-analysis

Federated learning for COVID-19 mortality prediction in a multicentric sample of 21 hospitals

We evaluated Federated Learning (FL) strategies for predicting COVID-19 mortality using a multicenter sample of 17,022 patients from 21 diverse Brazilian hospitals. We tested horizontal FL architectures employing Logistic Regression (LR) and a Multi-Layer Perceptron (MLP) via parameter aggregation, alongside a novel Federated Random Forest (RF) using ensemble aggregation. Performance gain ( Δ AUC, calculated as AUC   federated minus AUC   local ) was quantified using bootstrap analysis to determine 95% confidence intervals. FL models demonstrated a beneficial collaborative effect. The average Δ AUC across the network was +0.0018 for LR, +0.0599 for MLP, and +0.0528 for RF. Crucially, the gain’s magnitude and statistical significance showed a strong inverse correlation with local patient volume (N). Substantial and statistically significant gains concentrated in data-limited institutions (N < 500). For example, the smallest hospital (N=86) achieved a remarkable Δ AUC of 0.3682 (95% CI [0.0908, 0.6307]) with the RF model. However, interpreting these benefits requires caution because the 95% CIs for Δ AUC crossed zero for the majority of hospitals, suggesting the collaborative model’s statistical advantage is not universally certain at every site. This trade-off was particularly evident with the MLP model which, despite achieving the highest average Δ AUC, was the most volatile algorithm, registering the maximum performance degradation in the network ( Δ AUC = –0.0884, 95% CI [–0.1527, –0.0273]) due to its high sensitivity to local data distribution disparities (non-IID). This study validates FL as an equity-enabling mechanism that effectively enhances predictive capacity where local data scarcity is highest. Our findings underscore that maximizing the most statistically certain benefits of FL requires continuous monitoring and local validation for successful clinical deployment across diverse settings.

Dental services use prediction among adults in Southern Brazil: A gender and racial fairness-oriented machine learning approach

Strategies for detecting and mitigating dataset shift in machine learning for health predictions: A systematic review

Artificial intelligence for the diagnosis of erythematous-squamous dermatological diseases: technological contributions to primary care

Transfer learning for COVID-19 predictive modeling: A multicenter study of 12 hospitals

Development and evaluation of machine learning training strategies for neonatal mortality prediction using multicountry data

Global Health in the Age of AI: Charting a Course for Ethical Implementation and Societal Benefit

Predicting all-cause mortality with machine learning among Brazilians aged 50 and over: results from The Brazilian Longitudinal Study of Ageing (ELSI-Brazil)

Prediction of Hypertension in the Pediatric Population Using Machine Learning and Transfer Learning: A Multicentric Analysis of the SAYCARE Study

ObjectiveTo develop a machine learning (ML) model utilizing transfer learning (TL) techniques to predict hypertension in children and adolescents across South America.MethodsData from two cohorts (children and adolescents) in seven South American cities were analyzed. A TL strategy was implemented by transferring knowledge from a CatBoost model trained on the children’s sample and adapting it to the adolescent sample. Model performance was evaluated using standard metrics.ResultsAmong children, the prevalence of normal blood pressure was 88.9% (301 participants), while 14.1% (50 participants) had elevated blood pressure (EBP). In the adolescent group, the prevalence of normal blood pressure was 92.5% (284 participants), with 7.5% (23 participants) presenting with EBP. Random Forest, XGBoost, and LightGBM achieved high accuracy (0.90) for children, with XGBoost and LightGBM demonstrating superior recall (0.50) and AUC-ROC (0.74). For adolescents, models without TL showed poor performance, with accuracy and recall values remaining low and AUC-ROC ranging from 0.46 to 0.56. After applying TL, model performance improved significantly, with CatBoost achieving an AUC-ROC of 0.82, accuracy of 1.0, and recall of 0.18.ConclusionSoft drinks, filled cookies, and chips were key dietary predictors of elevated blood pressure, with higher intake in adolescents. Machine learning with transfer learning effectively identified these risks, emphasizing the need for early dietary interventions to prevent hypertension and support cardiovascular health in pediatric populations.

Multicenter comparative analysis of local and aggregated data training strategies in COVID-19 outcome prediction with Machine learning

Machine learning (ML) is a promising tool in assisting clinical decision-making for improving diagnosis and prognosis, especially in developing regions. It is often used with large samples, aggregating data from different regions and hospitals. However, it is unclear how this affects predictions in local centers. This study aims to compare data aggregation strategies of several hospitals in Brazil with a local training strategy in each hospital to predict two COVID-19 outcomes: Intensive Care Unit admission (ICU) and mechanical ventilation use (MV). The study included 6,046 patients from 14 hospitals, with local sample sizes ranging from 47 to 1500 patients. Machine learning models were trained using extreme gradient boosting, lightGBM, and catboost for structured data. Seven data aggregation strategies based on hospital geographic regions were compared with local training, and the best strategy was determined by analyzing the area under the ROC curve (AUROC). SHAP (Shapley Additive exPlanations) values were used to assess the contribution of variables to predictions. Additionally, a metafeatures analysis examined how hospital characteristics influence the selection of the best strategy. The study found that the local training strategy was the most effective approach, in the case of ICU outcomes, for 11 of the 14 hospitals (79%), and, in the case of MV, for 10 hospitals (71%). Metafeatures analysis suggested that hospitals with smaller sample sizes generally performed better using an aggregated data strategy compared to local training. Our study brings to light an important concern about the impact of grouping data from different hospitals in predictive machine learning models. These findings contribute to the ongoing debate about the trade-off between increasing sample size and bringing together heterogeneous scenarios.

Predictive modeling of gestational weight gain: a machine learning multiclass classification study

Does machine learning have a high performance to predict obesity among adults and older adults? A systematic review and meta-analysis

Machine learning for predicting Chagas disease infection in rural areas of Brazil

Introduction Chagas disease is a severe parasitic illness that is prevalent in Latin America and often goes unaddressed. Early detection and treatment are critical in preventing the progression of the illness and its associated life-threatening complications. In recent years, machine learning algorithms have emerged as powerful tools for disease prediction and diagnosis. Methods In this study, we developed machine learning algorithms to predict the risk of Chagas disease based on five general factors: age, gender, history of living in a mud or wooden house, history of being bitten by a triatomine bug, and family history of Chagas disease. We analyzed data from the Retrovirus Epidemiology Donor Study (REDS) to train five popular machine learning algorithms. The sample comprised 2,006 patients, divided into 75% for training and 25% for testing algorithm performance. We evaluated the model performance using precision, recall, and AUC-ROC metrics. Results The Adaboost algorithm yielded an AUC-ROC of 0.772, a precision of 0.199, and a recall of 0.612. We simulated the decision boundary using various thresholds and observed that in this dataset a threshold of 0.45 resulted in a 100% recall. This finding suggests that employing such a threshold could potentially save 22.5% of the cost associated with mass testing of Chagas disease. Conclusion Our findings highlight the potential of applying machine learning to improve the sensitivity and effectiveness of Chagas disease diagnosis and prevention. Furthermore, we emphasize the importance of integrating socio-demographic and environmental factors into neglected disease prediction models to enhance their performance.

Machine learning for longitudinal mortality risk prediction in patients with malignant neoplasm in São Paulo, Brazil

Fairness of Machine Learning Algorithms for Predicting Foregone Preventive Dental Care for Adults

ImportanceAccess to routine dental care prevents advanced dental disease and improves oral and overall health. Identifying individuals at risk of foregoing preventive dental care can direct prevention efforts toward high-risk populations.ObjectiveTo predict foregone preventive dental care among adults overall and in sociodemographic subgroups and to assess the algorithmic fairness.Design, Setting, and ParticipantsThis prognostic study was a secondary analyses of longitudinal data from the US Medical Expenditure Panel Survey (MEPS) from 2016 to 2019, each with 2 years of follow-up. Participants included adults aged 18 years and older. Data analysis was performed from December 2022 to June 2023.ExposureA total of 50 predictors, including demographic and socioeconomic characteristics, health conditions, behaviors, and health services use, were assessed.Main Outcomes and MeasuresThe outcome of interest was foregoing preventive dental care, defined as either cleaning, general examination, or an appointment with the dental hygienist, in the past year.ResultsAmong 32 234 participants, the mean (SD) age was 48.5 (18.2) years and 17 386 participants (53.9%) were female; 1935 participants (6.0%) were Asian, 5138 participants (15.9%) were Black, 7681 participants (23.8%) were Hispanic, 16 503 participants (51.2%) were White, and 977 participants (3.0%) identified as other (eg, American Indian and Alaska Native) or multiple racial or ethnic groups. There were 21 083 (65.4%) individuals who missed preventive dental care in the past year. The algorithms demonstrated high performance, achieving an area under the receiver operating characteristic curve (AUC) of 0.84 (95% CI, 0.84-0.85) in the overall population. While the full sample model performed similarly when applied to White individuals and older adults (AUC, 0.88; 95% CI, 0.87-0.90), there was a loss of performance for other subgroups. Removing the subgroup-sensitive predictors (ie, race and ethnicity, age, and income) did not impact model performance. Models stratified by race and ethnicity performed similarly or worse than the full model for all groups, with the lowest performance for individuals who identified as other or multiple racial groups (AUC, 0.76; 95% CI, 0.70-0.81). Previous pattern of dental visits, health care utilization, dental benefits, and sociodemographic characteristics were the highest contributing predictors to the models’ performance.Conclusions and RelevanceFindings of this prognostic study using cohort data suggest that tree-based ensemble machine learning models could accurately predict adults at risk of foregoing preventive dental care and demonstrated bias against underrepresented sociodemographic groups. These results highlight the importance of evaluating model fairness during development and testing to avoid exacerbating existing biases.

Inteligência artificial, noções básicas para os profissionais de saúde

Frente a importância que as tecnologias computacionais tem ganhando para a ciência em saúde, os autores propõem uma revisão narrativa prática para esclarecer e nortear conceitos em inteligência artificial. Nesse contexto, torna-se necessário entendermos as etapas que compõe a aprendizagem de máquina e serão discutidos ao longo dessa dissertação, entre eles a captação dos dados, limpeza e preparação dos dados, treino do modelo, otimização de hiperparametros, e disponibilização do modelo.   

Spatial Clusters of Cancer Mortality in Brazil: A Machine Learning Modeling Approach

Objectives: Our aim was to test if machine learning algorithms can predict cancer mortality (CM) at an ecological level and use these results to identify statistically significant spatial clusters of excess cancer mortality (eCM).Methods: Age-standardized CM was extracted from the official databases of Brazil. Predictive features included sociodemographic and health coverage variables. Machine learning algorithms were selected and trained with 70% of the data, and the performance was tested with the remaining 30%. Clusters of eCM were identified using SatScan. Additionally, separate analyses were performed for the 10 most frequent cancer types.Results: The gradient boosting trees algorithm presented the highest coefficient of determination (R2 = 0.66). For total cancer, all algorithms overlapped in the region of Bagé (27% eCM). For esophageal cancer, all algorithms overlapped in west Rio Grande do Sul (48%–96% eCM). The most significant cluster for stomach cancer was in Macapá (82% eCM). The most important variables were the percentage of the white population and residents with computers.Conclusion: We found consistent and well-defined geographic regions in Brazil with significantly higher than expected cancer mortality.

A Machine Learning Prediction Model for Immediate Graft Function After Deceased Donor Kidney Transplantation

Background. After kidney transplantation (KTx), the graft can evolve from excellent immediate graft function (IGF) to total absence of function requiring dialysis. Recipients with IGF do not seem to benefit from using machine perfusion, an expensive procedure, in the long term when compared with cold storage. This study proposes to develop a prediction model for IGF in KTx deceased donor patients using machine learning algorithms. Methods. Unsensitized recipients who received their first KTx deceased donor between January 1, 2010, and December 31, 2019, were classified according to the conduct of renal function after transplantation. Variables related to the donor, recipient, kidney preservation, and immunology were used. The patients were randomly divided into 2 groups: 70% were assigned to the training and 30% to the test group. Popular machine learning algorithms were used: eXtreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine, Gradient Boosting classifier, Logistic Regression, CatBoost classifier, AdaBoost classifier, and Random Forest classifier. Comparative performance analysis on the test dataset was performed using the results of the AUC values, sensitivity, specificity, positive predictive value, negative predictive value, and F1 score. Results. Of the 859 patients, 21.7% (n = 186) had IGF. The best predictive performance resulted from the eXtreme Gradient Boosting model (AUC, 0.78; 95% CI, 0.71–0.84; sensitivity, 0.64; specificity, 0.78). Five variables with the highest predictive value were identified. Conclusions. Our results indicated the possibility of creating a model for the prediction of IGF, enhancing the selection of patients who would benefit from an expensive treatment, as in the case of machine perfusion preservation.

Improving the performance of machine learning algorithms for health outcomes predictions in multicentric cohorts

Machine learning algorithms are being increasingly used in healthcare settings but their generalizability between different regions is still unknown. This study aims to identify the strategy that maximizes the predictive performance of identifying the risk of death by COVID-19 in different regions of a large and unequal country. This is a multicenter cohort study with data collected from patients with a positive RT-PCR test for COVID-19 from March to August 2020 (n = 8477) in 18 hospitals, covering all five Brazilian regions. Of all patients with a positive RT-PCR test during the period, 2356 (28%) died. Eight different strategies were used for training and evaluating the performance of three popular machine learning algorithms (extreme gradient boosting, lightGBM, and catboost). The strategies ranged from only using training data from a single hospital, up to aggregating patients by their geographic regions. The predictive performance of the algorithms was evaluated by the area under the ROC curve (AUROC) on the test set of each hospital. We found that the best overall predictive performances were obtained when using training data from the same hospital, which was the winning strategy for 11 (61%) of the 18 participating hospitals. In this study, the use of more patient data from other regions slightly decreased predictive performance. However, models trained in other hospitals still had acceptable performances and could be a solution while data for a specific hospital is being collected.

Data-driven decision making for the screening of cognitive impairment in primary care: a machine learning approach using data from the ELSA-Brasil study

Emergency department use and Artificial Intelligence in Pelotas: design and baseline results

Objetivo: To describe the initial baseline results of a population-based study, as well as a protocol in order to evaluate the performance of different machine learning algorithms with the objective of predicting the demand for urgent and emergency services in a representative sample of adults from the urban area of Pelotas, Southern Brazil. Methods: The study is entitled “Emergency department use and Artificial Intelligence in PELOTAS (RS) (EAI PELOTAS)” (https://wp.ufpel.edu.br/eaipelotas/). Between September and December 2021, a baseline was carried out with participants. A follow-up was planned to be conducted after 12 months in order to assess the use of urgent and emergency services in the last year. Afterwards, machine learning algorithms will be tested to predict the use of urgent and emergency services over one year. Results: In total, 5,722 participants answered the survey, mostly females (66.8%), with an average age of 50.3 years. The mean number of household people was 2.6. Most of the sample has white skin color and incomplete elementary school or less. Around 30% of the sample has obesity, 14% diabetes, and 39% hypertension. Conclusion: The present paper presented a protocol describing the steps that were and will be taken to produce a model capable of predicting the demand for urgent and emergency services in one year among residents of Pelotas, in Rio Grande do Sul state.

Physician preference for receiving machine learning predictive results: A cross-sectional multicentric study

Artificial intelligence (AI) algorithms are transforming several areas of the digital world and are increasingly being applied in healthcare. Mobile apps based on predictive machine learning models have the potential to improve health outcomes, but there is still no consensus on how to inform doctors about their results. The aim of this study was to investigate how healthcare professionals prefer to receive predictions generated by machine learning algorithms. A systematic search in MEDLINE, via PubMed, EMBASE and Web of Science was first performed. We developed a mobile app, RandomIA, to predict the occurrence of clinical outcomes, initially for COVID-19 and later expected to be expanded to other diseases. A questionnaire called System Usability Scale (SUS) was selected to assess the usability of the mobile app. A total of 69 doctors from the five regions of Brazil tested RandomIA and evaluated three different ways to visualize the predictions. For prognostic outcomes (mechanical ventilation, admission to an intensive care unit, and death), most doctors (62.9%) preferred a more complex visualization, represented by a bar graph with three categories (low, medium, and high probability) and a probability density graph for each outcome. For the diagnostic prediction of COVID-19, there was also a majority preference (65.4%) for the same option. Our results indicate that doctors could be more inclined to prefer receiving detailed results from predictive machine learning algorithms.

Clusters of Pregnant Women with Severe Acute Respiratory Syndrome Due to COVID-19: An Unsupervised Learning Approach

COVID-19 has been widely explored in relation to its symptoms, outcomes, and risk profiles for the severe form of the disease. Our aim was to identify clusters of pregnant and postpartum women with severe acute respiratory syndrome (SARS) due to COVID-19 by analyzing data available in the Influenza Epidemiological Surveillance Information System of Brazil (SIVEP-Gripe) between March 2020 and August 2021. The study’s population comprised 16,409 women aged between 10 and 49 years old. Multiple correspondence analyses were performed to summarize information from 28 variables related to symptoms, comorbidities, and hospital characteristics into a set of continuous principal components (PCs). The population was segmented into three clusters based on an agglomerative hierarchical cluster analysis applied to the first 10 PCs. Cluster 1 had a higher frequency of younger women without comorbidities and with flu-like symptoms; cluster 2 was represented by women who reported mainly ageusia and anosmia; cluster 3 grouped older women with the highest frequencies of comorbidities and poor outcomes. The defined clusters revealed different levels of disease severity, which can contribute to the initial risk assessment of the patient, assisting the referral of these women to health services with an appropriate level of complexity.

Machine Learning for Hypertension Prediction: a Systematic Review

Early identification of older individuals at risk of mobility decline with machine learning

Machine learning for predicting chronic diseases: a systematic review

Neonatal mortality prediction with routinely collected data: a machine learning approach

Background Recent decreases in neonatal mortality have been slower than expected for most countries. This study aims to predict the risk of neonatal mortality using only data routinely available from birth records in the largest city of the Americas. Methods A probabilistic linkage of every birth record occurring in the municipality of São Paulo, Brazil, between 2012 e 2017 was performed with the death records from 2012 to 2018 (1,202,843 births and 447,687 deaths), and a total of 7282 neonatal deaths were identified (a neonatal mortality rate of 6.46 per 1000 live births). Births from 2012 and 2016 (N = 941,308; or 83.44% of the total) were used to train five different machine learning algorithms, while births occurring in 2017 (N = 186,854; or 16.56% of the total) were used to test their predictive performance on new unseen data. Results The best performance was obtained by the extreme gradient boosting trees (XGBoost) algorithm, with a very high AUC of 0.97 and F1-score of 0.55. The 5% births with the highest predicted risk of neonatal death included more than 90% of the actual neonatal deaths. On the other hand, there were no deaths among the 5% births with the lowest predicted risk. There were no significant differences in predictive performance for vulnerable subgroups. The use of a smaller number of variables (WHO’s five minimum perinatal indicators) decreased overall performance but the results still remained high (AUC of 0.91). With the addition of only three more variables, we achieved the same predictive performance (AUC of 0.97) as using all the 23 variables originally available from the Brazilian birth records. Conclusion Machine learning algorithms were able to identify with very high predictive performance the neonatal mortality risk of newborns using only routinely collected data.

Predictors of tooth loss: A machine learning approach

Introduction Little is understood about the socioeconomic predictors of tooth loss, a condition that can negatively impact individual’s quality of life. The goal of this study is to develop a machine-learning algorithm to predict complete and incremental tooth loss among adults and to compare the predictive performance of these models. Methods We used data from the National Health and Nutrition Examination Survey from 2011 to 2014. We developed multiple machine-learning algorithms and assessed their predictive performances by examining the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, and positive and negative predictive values. Results The extreme gradient boosting trees presented the highest performance in the prediction of edentulism (AUC = 88.7%; 95%CI: 87.1, 90.2), the absence of a functional dentition (AUC = 88.3% 95%CI: 87.3,89.3) and for predicting missing any tooth (AUC = 83.2%; 95%CI, 82.0, 84.4). Although, as expected, age and routine dental care emerged as strong predictors of tooth loss, the machine learning approach identified additional predictors, including socioeconomic conditions. Indeed, the performance of models incorporating socioeconomic characteristics was better at predicting tooth loss than those relying on clinical dental indicators alone. Conclusions Future application of machine-learning algorithm, with longitudinal cohorts, for identification of individuals at risk for tooth loss could assist clinicians to prioritize interventions directed toward the prevention of tooth loss.

Predição de absenteísmo docente na rede pública com machine learning

OBJETIVO Predizer o risco de ausência laboral decorrente de morbidades dos docentes que atuam na educação infantil na rede pública municipal, com o uso de algoritmos de machine learning. MÉTODOS Trata-se de um estudo transversal utilizando dados secundários, públicos e anônimos da Relação Anual de Informações Sociais, selecionando professores da educação infantil que atuaram na rede pública municipal do estado de São Paulo entre 2014 e 2018 (n = 174.294). Foram também vinculados dados da média de alunos por turma e número de habitantes no município. Os dados foram separados em treinamento e teste, utilizando os registros de 2014 a 2016 (n = 103.357) para treinar cinco modelos preditivos e os dados de 2017 a 2018 (n = 70.937) para testar seus desempenhos em dados novos. A performance preditiva dos algoritmos foi avaliada por meio do valor da área abaixo da curva ROC (AUROC). RESULTADOS Todos os cinco algoritmos testados apresentaram área abaixo da curva acima de 0,76. O algoritmo com melhor performance preditiva (redes neurais artificiais) obteve 0,79 de área abaixo da curva, com acurácia de 71,52%, sensibilidade de 72,86%, especificidade de 70,52% e kappa de 0,427 nos dados de teste. CONCLUSÃO É possível predizer casos de afastamentos por morbidade em docentes da rede pública com machine learning usando dados públicos. O melhor algoritmo apresentou melhor resultado da área abaixo da curva quando comparado ao modelo de referência (regressão logística). Os algoritmos podem contribuir para predições mais assertivas na área da saúde pública e da saúde do trabalhador, permitindo acompanhar e ajudar a prevenir afastamentos por morbidade desses trabalhadores.

Cause-specific mortality prediction in older residents of São Paulo, Brazil: a machine learning approach

Background Populational ageing has been increasing in a remarkable rate in developing countries. In this scenario, preventive strategies could help to decrease the burden of higher demands for healthcare services. Machine learning algorithms have been increasingly applied for identifying priority candidates for preventive actions, presenting a better predictive performance than traditional parsimonious models. Methods Data were collected from the Health, Well Being and Aging (SABE) Study, a representative sample of older residents of São Paulo, Brazil. Machine learning algorithms were applied to predict death by diseases of respiratory system (DRS), diseases of circulatory system (DCS), neoplasms and other specific causes within 5 years, using socioeconomic, demographic and health features. The algorithms were trained in a random sample of 70% of subjects, and then tested in the other 30% unseen data. Results The outcome with highest predictive performance was death by DRS (AUC−ROC = 0.89), followed by the other specific causes (AUC−ROC = 0.87), DCS (AUC−ROC = 0.67) and neoplasms (AUC−ROC = 0.52). Among only the 25% of individuals with the highest predicted risk of mortality from DRS were included 100% of the actual cases. The machine learning algorithms with the highest predictive performance were light gradient boosted machine and extreme gradient boosting. Conclusion The algorithms had a high predictive performance for DRS, but lower for DCS and neoplasms. Mortality prediction with machine learning can improve clinical decisions especially regarding targeted preventive measures for older individuals.

Machine learning and national health data to improve evidence: Finding segmentation in individuals without private insurance

Data Leakage in Health Outcomes Prediction With Machine Learning. Comment on “Prediction of Incident Hypertension Within the Next Year: Prospective Study Using Statewide Electronic Health Records and Machine Learning”

A multipurpose machine learning approach to predict COVID-19 negative prognosis in São Paulo, Brazil

The new coronavirus disease (COVID-19) is a challenge for clinical decision-making and the effective allocation of healthcare resources. An accurate prognostic assessment is necessary to improve survival of patients, especially in developing countries. This study proposes to predict the risk of developing critical conditions in COVID-19 patients by training multipurpose algorithms. We followed a total of 1040 patients with a positive RT-PCR diagnosis for COVID-19 from a large hospital from São Paulo, Brazil, from March to June 2020, of which 288 (28%) presented a severe prognosis, i.e. Intensive Care Unit (ICU) admission, use of mechanical ventilation or death. We used routinely-collected laboratory, clinical and demographic data to train five machine learning algorithms (artificial neural networks, extra trees, random forests, catboost, and extreme gradient boosting). We used a random sample of 70% of patients to train the algorithms and 30% were left for performance assessment, simulating new unseen data. In order to assess if the algorithms could capture general severe prognostic patterns, each model was trained by combining two out of three outcomes to predict the other. All algorithms presented very high predictive performance (average AUROC of 0.92, sensitivity of 0.92, and specificity of 0.82). The three most important variables for the multipurpose algorithms were ratio of lymphocyte per C-reactive protein, C-reactive protein and Braden Scale. The results highlight the possibility that machine learning algorithms are able to predict unspecific negative COVID-19 outcomes from routinely-collected data.

Machine learning analysis to predict health outcomes among emergency department users in Southern Brazil: a protocol study

: Objective: Emergency services are essential to the organization of the health care system. Nevertheless, they face different operational difficulties, including overcrowded services, largely explained by their inappropriate use and the repeated visits from users. Although a known situation, information on the theme is scarce in Brazil, particularly regarding longitudinal user monitoring. Thus, this project aims to evaluate the predictive performance of different machine learning algorithms to estimate the inappropriate and repeated use of emergency services and mortality. Methods: To that end, a study will be conducted in the municipality of Pelotas, Rio Grande do Sul, with around five thousand users of the municipal emergency department. Results: If the study is successful, we will provide an algorithm that could be used in clinical practice to assist health professionals in decision-making within hospitals. Different knowledge dissemination strategies will be used to increase the capacity of the study to produce innovations for the organization of the health system and services. Conclusion: A high performance predictive model may be able to help decisionmaking in the emergency services, improving quality of care.

Machine learning to predict 30-day quality-adjusted survival in critically ill patients with cancer

Machine learning para análises preditivas em saúde: exemplo de aplicação para predizer óbito em idosos de São Paulo, Brasil

Este estudo objetiva apresentar as etapas relacionadas à utilização de algoritmos de machine learning para análises preditivas em saúde. Para isso, foi realizada uma aplicação com base em dados de idosos residentes no Município de São Paulo, Brasil, participantes do estudo Saúde Bem-estar e Envelhecimento (SABE) (n = 2.808). A variável resposta foi representada pela ocorrência de óbito em até cinco anos após o ingresso do idoso no estudo (n = 423), e os preditores, por 37 variáveis relacionadas ao perfil demográfico, socioeconômico e de saúde do idoso. A aplicação foi organizada de acordo com as seguintes etapas: divisão dos dados em treinamento (70%) e teste (30%), pré-processamento dos preditores, aprendizado e avaliação de modelos. Na etapa de aprendizado, foram utilizados cinco algoritmos para o ajuste de modelos: regressão logística com e sem penalização, redes neurais, gradient boosted trees e random forest. Os hiperparâmetros dos algoritmos foram otimizados por validação cruzada 10-fold, para selecionar aqueles correspondentes aos melhores modelos. Para cada algoritmo, o melhor modelo foi avaliado em dados de teste por meio da área abaixo da curva (AUC) ROC e medidas relacionadas. Todos os modelos apresentaram AUC ROC superior a 0,70. Para os três modelos com maior AUC ROC (redes neurais e regressão logística com penalização de lasso e sem penalização, respectivamente), foram também avaliadas medidas de qualidade da probabilidade predita. Espera-se que, com o aumento da disponibilidade de dados e de capital humano capacitado, seja possível desenvolver modelos preditivos de machine learning com potencial para auxiliar profissionais de saúde na tomada de melhores decisões.

Perspectivas do uso de mineração de dados e aprendizado de máquina em saúde e segurança no trabalho

Introdução: a variedade, volume e velocidade de geração de dados (big data) possibilitam novas e mais complexas análises. Objetivo: discutir e apresentar técnicas de mineração de dados (data mining) e de aprendizado de máquina (machine learning) para auxiliar pesquisadores de Saúde e Segurança no Trabalho (SST) na escolha da técnica adequada para lidar com big data. Métodos: revisão bibliográfica com foco em data mining e no uso de análises preditivas com machine learning e suas aplicações para auxiliar diagnósticos e predição de riscos em SST. Resultados: a literatura indica que aplicações de data mining com algoritmos de machine learning para análises preditivas em saúde pública e em SST apresentam melhor desempenho em comparação com análises tradicionais. São sugeridas técnicas de acordo com o tipo de pesquisa almejada. Discussão: data mining tem se tornado uma alternativa cada vez mais comum para lidar com bancos de dados de saúde pública, possibilitando analisar grandes volumes de dados de morbidade e mortalidade. Tais técnicas não visam substituir o fator humano, mas auxiliar em processos de tomada de decisão, servir de ferramenta para a análise estatística e gerar conhecimento para subsidiar ações que possam melhorar a qualidade de vida do trabalhador.

Overachieving Municipalities in Public Health: A Machine-learning Approach

Background: Identifying successful public health ideas and practices is a difficult challenge towing to the presence of complex baseline characteristics that can affect health outcomes. We propose the use of machine learning algorithms to predict life expectancy at birth, and then compare health-related characteristics of the under- and overachievers (i.e., municipalities that have a worse and better outcome than predicted, respectively). Methods: Our outcome was life expectancy at birth for Brazilian municipalities, and we used as predictors 60 local characteristics that are not directly controlled by public health officials (e.g., socioeconomic factors). Results: The highest predictive performance was achieved by an ensemble of machine learning algorithms (cross-validated mean squared error of 0.168), including a 35% gain in comparison with standard decision trees. Overachievers presented better results regarding primary health care, such as higher coverage of the massive multidisciplinary program Family Health Strategy. On the other hand, underachievers performed more cesarean deliveries and mammographies and had more life-support health equipment. Conclusions: The findings suggest that analyzing the predicted value of a health outcome may bring insights about good public health practices.