Machine learning-based model for prediction of clinical deterioration in hospitalized patients by COVID 19

Abstract

Despite the publication of great number of tools to aid decisions in COVID-19 patients, there is a lack of good instruments to predict clinical deterioration. COVID19-Osakidetza is a prospective cohort study recruiting COVID-19 patients. We collected information from baseline to discharge on: sociodemographic characteristics, comorbidities and associated medications, vital signs, treatment received and lab test results. Outcome was need for intensive ventilatory support (with at least standard high-flow oxygen face mask with a reservoir bag for at least 6 h and need for more intensive therapy afterwards or Optiflow high-flow nasal cannula or noninvasive or invasive mechanical ventilation) and/or admission to a critical care unit and/or death during hospitalization. We developed a Catboost model summarizing the findings using Shapley Additive Explanations. Performance of the model was assessed using area under the receiver operating characteristic and prediction recall curves (AUROC and AUPRC respectively) and calibrated using the Hosmer-Lemeshow test. Overall, 1568 patients were included in the derivation cohort and 956 in the (external) validation cohort. The percentages of patients who reached the composite endpoint were 23.3% vs 20% respectively. The strongest predictors of clinical deterioration were arterial blood oxygen pressure, followed by age, levels of several markers of inflammation (procalcitonin, LDH, CRP) and alterations in blood count and coagulation. Some medications, namely, ATC AO2 (antiacids) and N05 (neuroleptics) were also among the group of main predictors, together with C03 (diuretics). In the validation set, the CatBoost AUROC was 0.79, AUPRC 0.21 and Hosmer-Lemeshow test statistic 0.36. We present a machine learning-based prediction model with excellent performance properties to implement in EHRs. Our main goal was to predict progression to a score of 5 or higher on the WHO Clinical Progression Scale before patients required mechanical ventilation. Future steps are to externally validate the model in other settings and in a cohort from a different period and to apply the algorithm in clinical practice.Registration: ClinicalTrials.gov Identifier: NCT04463706.

Conflict of interest statement

The authors declare no competing interests.

© 2022. The Author(s).

Figures

Figure 1

Flow chart.

Figure 2

Main predictors in catboost model in (a) derivation-internal validation and (b) external validation datasets. PO2-A partial arterial oxygen concentration, PCR C-reactive proteine, PCT procalcitonine, Edad age, LDH lactate dehydrogenase, PLT platelets, ADE RED blood cell distribution width, CREA creatinine, Mon%A Total count of monocytes, CK creatine kinase, Dimer D dimer, EOS%A Percentage of eoshinophils, AO2_12 Antacids in the last 12 months, N05_12 neuroleptics in the last 12 months, MON#A total count of monocytes, GLU glucose, EOS#A TOTAL count of eosynophils, lin#A total count of lynphocites, Neu%A percentage of neutrophils, CO3_12 diuretics in the last 12 months.

Figure 3

Predictive performance of the catboost model in (a) derivation and (b) validation sets.

Figure 4

Calibration performance in (a) derivation and (b) validation sets.

Figure 5

Differences between traditional development and machine-learning based prediction models.