Using data-driven rules to predict mortality in severe community acquired pneumonia

Machine learning-derived prediction of in-hospital mortality in patients with severe acute respiratory infection: analysis of claims data from the German-wide Helios hospital network

BACKGROUND

Severe acute respiratory infections (SARI) are the most common infectious causes of death. Previous work regarding mortality prediction models for SARI using machine learning (ML) algorithms that can be useful for both individual risk stratification and quality of care assessment is scarce. We aimed to develop reliable models for mortality prediction in SARI patients utilizing ML algorithms and compare its performances with a classic regression analysis approach.

METHODS

Administrative data (dataset randomly split 75%/25% for model training/testing) from years 2016-2019 of 86 German Helios hospitals was retrospectively analyzed. Inpatient SARI cases were defined by ICD-codes J09-J22. Three ML algorithms were evaluated and its performance compared to generalized linear models (GLM) by computing receiver operating characteristic area under the curve (AUC) and area under the precision-recall curve (AUPRC).

RESULTS

The dataset contained 241,988 inpatient SARI cases (75 years or older: 49%; male 56.2%). In-hospital mortality was 11.6%. AUC and AUPRC in the testing dataset were 0.83 and 0.372 for GLM, 0.831 and 0.384 for random forest (RF), 0.834 and 0.382 for single layer neural network (NNET) and 0.834 and 0.389 for extreme gradient boosting (XGBoost). Statistical comparison of ROC AUCs revealed a better performance of NNET and XGBoost as compared to GLM.

CONCLUSION

ML algorithms for predicting in-hospital mortality were trained and tested on a large real-world administrative dataset of SARI patients and showed good discriminatory performances. Broad application of our models in clinical routine practice can contribute to patients' risk assessment and quality management.

Comparison of deep learning approaches to predict COVID-19 infection

The SARS-CoV2 virus, which causes COVID-19 (coronavirus disease) has become a pandemic and has expanded all over the world. Because of increasing number of cases day by day, it takes time to interpret the laboratory findings thus the limitations in terms of both treatment and findings are emerged. Due to such limitations, the need for clinical decisions making system with predictive algorithms has arisen. Predictive algorithms could potentially ease the strain on healthcare systems by identifying the diseases. In this study, we perform clinical predictive models that estimate, using deep learning and laboratory data, which patients are likely to receive a COVID-19 disease. To evaluate the predictive performance of our models, precision, F1-score, recall, AUC, and accuracy scores calculated. Models were tested with 18 laboratory findings from 600 patients and validated with 10 fold cross-validation and train-test split approaches. The experimental results indicate that our predictive models identify patients that have COVID-19 disease at an accuracy of 86.66%, F1-score of 91.89%, precision of 86.75%, recall of 99.42%, and AUC of 62.50%. It is observed that predictive models trained on laboratory findings could be used to predict COVID-19 infection, and can be helpful for medical experts to prioritize the resources correctly. Our models (available at (https://github.com/burakalakuss/COVID-19-Clinical)) can be employed to assists medical experts in validating their initial laboratory findings, and can also be used for clinical prediction studies.

Clinical Predictive Models for COVID-19: Systematic Study

BACKGROUND

COVID-19 is a rapidly emerging respiratory disease caused by SARS-CoV-2. Due to the rapid human-to-human transmission of SARS-CoV-2, many health care systems are at risk of exceeding their health care capacities, in particular in terms of SARS-CoV-2 tests, hospital and intensive care unit (ICU) beds, and mechanical ventilators. Predictive algorithms could potentially ease the strain on health care systems by identifying those who are most likely to receive a positive SARS-CoV-2 test, be hospitalized, or admitted to the ICU.

OBJECTIVE

The aim of this study is to develop, study, and evaluate clinical predictive models that estimate, using machine learning and based on routinely collected clinical data, which patients are likely to receive a positive SARS-CoV-2 test or require hospitalization or intensive care.

METHODS

Using a systematic approach to model development and optimization, we trained and compared various types of machine learning models, including logistic regression, neural networks, support vector machines, random forests, and gradient boosting. To evaluate the developed models, we performed a retrospective evaluation on demographic, clinical, and blood analysis data from a cohort of 5644 patients. In addition, we determined which clinical features were predictive to what degree for each of the aforementioned clinical tasks using causal explanations.

RESULTS

Our experimental results indicate that our predictive models identified patients that test positive for SARS-CoV-2 a priori at a sensitivity of 75% (95% CI 67%-81%) and a specificity of 49% (95% CI 46%-51%), patients who are SARS-CoV-2 positive that require hospitalization with 0.92 area under the receiver operator characteristic curve (AUC; 95% CI 0.81-0.98), and patients who are SARS-CoV-2 positive that require critical care with 0.98 AUC (95% CI 0.95-1.00).

CONCLUSIONS

Our results indicate that predictive models trained on routinely collected clinical data could be used to predict clinical pathways for COVID-19 and, therefore, help inform care and prioritize resources.

An exploration of mortality risk factors in non-severe pneumonia in children using clinical data from Kenya

BACKGROUND

Childhood pneumonia is the leading infectious cause of mortality in children younger than 5 years old. Recent updates to World Health Organization pneumonia guidelines recommend outpatient care for a population of children previously classified as high risk. This revision has been challenged by policymakers in Africa, where mortality related to pneumonia is higher than in other regions and often complicated by comorbidities. This study aimed to identify factors that best discriminate inpatient mortality risk in non-severe pneumonia and explore whether these factors offer any added benefit over the current criteria used to identify children with pneumonia requiring inpatient care.

METHODS

We undertook a retrospective cohort study of children aged 2-59 months admitted with a clinical diagnosis of pneumonia at 14 public hospitals in Kenya between February 2014 and February 2016. Using machine learning techniques, we analysed whether clinical characteristics and common comorbidities increased the risk of inpatient mortality for non-severe pneumonia. The topmost risk factors were subjected to decision curve analysis to explore if using them as admission criteria had any net benefit above the current criteria.

RESULTS

Out of 16,162 children admitted with pneumonia during the study period, 10,687 were eligible for subsequent analysis. Inpatient mortality within this non-severe group was 252/10,687 (2.36%). Models demonstrated moderately good performance; the partial least squares discriminant analysis model had higher sensitivity for predicting mortality in comparison to logistic regression. Elevated respiratory rate (≥70 bpm), age 2-11 months and weight-for-age Z-score (WAZ) < -3SD were highly discriminative of mortality. These factors ranked consistently across the different models. For a risk threshold probability of 7-14%, there is a net benefit to admitting the patient sub-populations with these features as additional criteria alongside those currently used to classify severe pneumonia. Of the population studied, 70.54% met at least one of these criteria. Sensitivity analyses indicated that the overall results were not significantly affected by variations in pneumonia severity classification criteria.

CONCLUSIONS

Children with non-severe pneumonia aged 2-11 months or with respiratory rate ≥ 70 bpm or very low WAZ experience risks of inpatient mortality comparable to severe pneumonia. Inpatient care is warranted in these high-risk groups of children.

The power of data mining in diagnosis of childhood pneumonia

Childhood pneumonia is the leading cause of death of children under the age of 5 years globally. Diagnostic information on the presence of infection, severity and aetiology (bacterial versus viral) is crucial for appropriate treatment. However, the derivation of such information requires advanced equipment (such as X-rays) and clinical expertise to correctly assess observational clinical signs (such as chest indrawing); both of these are often unavailable in resource-constrained settings. In this study, these challenges were addressed through the development of a suite of data mining tools, facilitating automated diagnosis through quantifiable features. Findings were validated on a large dataset comprising 780 children diagnosed with pneumonia and 801 age-matched healthy controls. Pneumonia was identified via four quantifiable vital signs (98.2% sensitivity and 97.6% specificity). Moreover, it was shown that severity can be determined through a combination of three vital signs and two lung sounds (72.4% sensitivity and 82.2% specificity); addition of a conventional biomarker (C-reactive protein) further improved severity predictions (89.1% sensitivity and 81.3% specificity). Finally, we demonstrated that aetiology can be determined using three vital signs and a newly proposed biomarker (lipocalin-2) (81.8% sensitivity and 90.6% specificity). These results suggest that a suite of carefully designed machine learning tools can be used to support multi-faceted diagnosis of childhood pneumonia in resource-constrained settings, compensating for the shortage of expensive equipment and highly trained clinicians.

Learning temporal rules to forecast instability in continuously monitored patients

Inductive machine learning, and in particular extraction of association rules from data, has been successfully used in multiple application domains, such as market basket analysis, disease prognosis, fraud detection, and protein sequencing. The appeal of rule extraction techniques stems from their ability to handle intricate problems yet produce models based on rules that can be comprehended by humans, and are therefore more transparent. Human comprehension is a factor that may improve adoption and use of data-driven decision support systems clinically via face validity. In this work, we explore whether we can reliably and informatively forecast cardiorespiratory instability (CRI) in step-down unit (SDU) patients utilizing data from continuous monitoring of physiologic vital sign (VS) measurements. We use a temporal association rule extraction technique in conjunction with a rule fusion protocol to learn how to forecast CRI in continuously monitored patients. We detail our approach and present and discuss encouraging empirical results obtained using continuous multivariate VS data from the bedside monitors of 297 SDU patients spanning 29 346 hours (3.35 patient-years) of observation. We present example rules that have been learned from data to illustrate potential benefits of comprehensibility of the extracted models, and we analyze the empirical utility of each VS as a potential leading indicator of an impending CRI event.

Community-acquired pneumonia (CAP) hospitalizations and deaths: is there a role for quality improvement through inter-hospital comparisons?

OBJECTIVE

To assess between-hospital variations in standardized in-hospital mortality ratios of community-acquired pneumonia (CAP), and identify possible leads for quality improvement.

DESIGN

We used an administrative database to estimate standardized in-hospital mortality ratios for 111 Belgian hospitals, by carrying out a set of hierarchical logistic regression models, intended to disentangle therapeutic attitudes and biases. To facilitate the detection of false-negative/positive results, we added an inconclusive zone to the funnel plots, derived from the results of the study. Data quality was validated by comparison with (i) alternative data from the largest Belgian Sickness Fund, (ii) published German hospital data and (iii) the results of an on-site audit.

SETTING

All Belgian hospital discharge records from 2004 to 2007.

STUDY PARTICIPANTS

A total of 111 776 adult patients were admitted for CAP.

MAIN OUTCOME MEASURE

Risk-adjusted standardized in-hospital mortality ratios.

RESULTS

Out of the 111 hospitals, we identified five and six outlying hospitals, with standardized mortality ratios of CAP consistently on the extremes of the distribution, as providing possibly better or worse care, respectively, and 18 other hospitals as having possible quality weaknesses/strengths. At the individuals' level of the analysis, adjusted odds ratios showed the paramount importance of old age, comorbidity and mechanical ventilation. The data compared well with the different validation sources.

CONCLUSIONS

Despite the limitations inherent to administrative data, it seemed possible to establish inter-hospital differences in standardized in-hospital mortality ratios of CAP and to identify leads for quality improvement. Monitoring is needed to assess progress in quality.

How Many Political Parties Should Brazil Have? A Data-Driven Method to Assess and Reduce Fragmentation in Multi-Party Political Systems

In June 2013, Brazil faced the largest and most significant mass protests in a generation. These were exacerbated by the population’s disenchantment towards its highly fragmented party system, which is composed by a very large number of political parties. Under these circumstances, presidents are constrained by informal coalition governments, bringing very harmful consequences to the country. In this work I propose ARRANGE, a dAta dRiven method foRAssessing and reduciNGparty fragmEntation in a country. ARRANGE uses as input the roll call data for congress votes on bills and amendments as a proxy for political preferences and ideology. With that, ARRANGE finds the minimum number of parties required to house all congressmen without decreasing party discipline. When applied to Brazil’s historical roll call data, ARRANGE was able to generate 23 distinct configurations that, compared with the status quo, have (i) a significant smaller number of parties, (ii) a higher discipline of partisans towards their parties and (iii) a more even distribution of partisans into parties. ARRANGE is fast and parsimonious, relying on a single, intuitive parameter.