Skip to main content
PLOS ONE logoLink to PLOS ONE
. 2020 Nov 12;15(11):e0239172. doi: 10.1371/journal.pone.0239172

Machine learning in predicting respiratory failure in patients with COVID-19 pneumonia—Challenges, strengths, and opportunities in a global health emergency

Davide Ferrari 1,2, Jovana Milic 1,3, Roberto Tonelli 3,4, Francesco Ghinelli 2, Marianna Meschiari 5, Sara Volpi 5, Matteo Faltoni 4, Giacomo Franceschi 5, Vittorio Iadisernia 5, Dina Yaacoub 5, Giacomo Ciusa 5, Erica Bacca 5, Carlotta Rogati 5, Marco Tutone 5, Giulia Burastero 5, Alessandro Raimondi 5, Marianna Menozzi 5, Erica Franceschini 5, Gianluca Cuomo 5, Luca Corradi 5, Gabriella Orlando 5, Antonella Santoro 5, Margherita Digaetano 5, Cinzia Puzzolante 5, Federica Carli 5, Vanni Borghi 5, Andrea Bedini 5, Riccardo Fantini 4, Luca Tabbì 4, Ivana Castaniere 3,4, Stefano Busani 6, Enrico Clini 4,7, Massimo Girardis 1,6, Mario Sarti 8, Andrea Cossarizza 7, Cristina Mussini 1,5, Federica Mandreoli 2, Paolo Missier 9,, Giovanni Guaraldi 1,5,‡,*
Editor: Paola Faverio10
PMCID: PMC7660476  PMID: 33180787

Abstract

Aims

The aim of this study was to estimate a 48 hour prediction of moderate to severe respiratory failure, requiring mechanical ventilation, in hospitalized patients with COVID-19 pneumonia.

Methods

This was an observational prospective study that comprised consecutive patients with COVID-19 pneumonia admitted to hospital from 21 February to 6 April 2020. The patients’ medical history, demographic, epidemiologic and clinical data were collected in an electronic patient chart. The dataset was used to train predictive models using an established machine learning framework leveraging a hybrid approach where clinical expertise is applied alongside a data-driven analysis. The study outcome was the onset of moderate to severe respiratory failure defined as PaO2/FiO2 ratio <150 mmHg in at least one of two consecutive arterial blood gas analyses in the following 48 hours. Shapley Additive exPlanations values were used to quantify the positive or negative impact of each variable included in each model on the predicted outcome.

Results

A total of 198 patients contributed to generate 1068 usable observations which allowed to build 3 predictive models based respectively on 31-variables signs and symptoms, 39-variables laboratory biomarkers and 91-variables as a composition of the two. A fourth “boosted mixed model” included 20 variables was selected from the model 3, achieved the best predictive performance (AUC = 0.84) without worsening the FN rate. Its clinical performance was applied in a narrative case report as an example.

Conclusion

This study developed a machine model with 84% prediction accuracy, which is able to assist clinicians in decision making process and contribute to develop new analytics to improve care at high technology readiness levels.

Background

COVID-19 pandemic found all health care system inadequately prepared and urges the need for new tools to face this unprecedented public health and clinical emergency. The clinical complexity of COVID-19 ranges from asymptomatic cases to severe pneumonia [13] whose progression to respiratory failure is difficult to predict. Pneumonia mostly occurs in the second or third week of a symptomatic infection and it is characterized by a mortality rate of 3–10%, which increases risk of multiorgan failure and mechanical ventilation [4]. Patients most commonly report the sudden onset of dyspnea during daily activities or rest. Prominent clinical signs include respiratory rate ≥ 30 breaths per minute, blood oxygen saturation ≤ 93%, partial pressure of arterial oxygen to fraction of inspired oxygen ratio (PaO2/FiO2) < 300 mmHg. This is an initial phase of acute respiratory distress syndrome (ARDS) that progressively leads to moderate to severe respiratory failure [4].

Overall, there is a high degree of uncertainty both in the progression of the patient’s health status and in the speed at which patients develop respiratory failure requiring mechanical ventilation.

Machine learning methods such as those employed to create the model have shown potential to produce predictive models that can be applied to assist and improve clinical decisions for a broad variety of outcomes [5, 6], and have recently been used in response to the COVID-19 emergency [79].

However, the most comprehensive review to date [9] finds that some risk prediction models attempt to predict risk of intensive care unit admission, ventilation, intubation, but most of these studies have shortcomings (high bias, poor reporting) that make them unsuitable for clinical decision-making. In contrast, the models presented in this work are explainable, meaning that they provide an easily understandable grounding for the choice of predictors and their relative importance on individual outcomes.

The aim of this study was to have a 48-hour prediction of moderate to severe respiratory failure, requiring mechanical ventilation, in hospitalized patients with COVID-19 pneumonia.

Methods

Study design

This observational prospective single center study included consecutive adult patients (≥18 years) admitted to Infectious Disease Clinic of the University Hospital of Modena, Italy from 21 February to 6 April 2020 with radiologically findings suggestive for COVID-19 pneumonia and confirmed by PCR method on nasopharyngeal swab.

All patients received treatment according to the Italian Society of Infectious Diseases’ Guidelines (SIMIT) recommendations [10] including oxygen supply to target SaO2 > 90%; hydroxychloroquine with or without azithromycin, and low molecular weight heparin. Lopinavir/ritonavir or darunavir/cobicistat was also used up to 18 March, when a clinical trial on the former did not show any benefit of protease inhibitors against the standard of care [11].

The study outcome was the onset of moderate to severe respiratory failure defined as PaO2/FiO2 ratio < 150 mmHg (≤ 13.3 kPa) in at least one of two consecutive arterial blood gas analyses in the following 48 hours.

Data source

Hospitalized patients with COVID-19 pneumonia were included if they had at least two arterial blood gas analyses measurements in the following 48 hours.

The patients’ full medical history including chronic comorbidities, demographic and epidemiological data were obtained at the hospital admission. Clinical data with signs and symptoms and complete blood count, coagulation, inflammatory and biochemical markers were routinely collected in the electronic patient charts.

Given the strict time dependent outcome definition, out of the initial sample of 295 patients and 2,889 data points available, 198 patients contributed to generate 1068 valuable observations. In detail, 603 observations contributed to the definition of respiratory failure (PaO2/FiO2 < 150 mmHg) and 465 did not meet this definition.

In the data collection period, the dataset was growing daily with the average of 84 new records per day, with a mean of 10 new data points/patient. Each data point included a complex record of observations from multiple categories: (1) signs and symptoms, (2) blood biomarkers, (3) respiratory assessment with PaO2/FiO2, (4) history of comorbidities (available in a subset of 119 patients). Some variables were collected daily, and others were recorded upon clinical indications.

All patients provided verbal, not written, informed consent due to isolation precautions. The study was approved by Regional ethical committee of Emilia Romagna (Area Vasta Nord protocol 426/2020).

Prediction models / machine learning methods

To be considered viable for clinical use, a predictive model must not only be accurate, it must also be (1) parsimonious, that is, it must achieve its accuracy using the minimal number of variables; (2) robust to missing data, an important feature in clinical emergency setting where not all observations are complete at each assessment; (3) transparent, in the sense that the model reveals the relative importance of each variable for each prediction it makes, which may be different for different patients. This is particularly important in clinical settings as it enables healthcare professionals to interpret the pathophysiological relationships between variables, arguably resulting in an increased trust in the model’s predictions.

Finally, the model should (4) minimize the number of false negatives (FN), that is, the risk of under-estimating the severity of a patient’s condition.

To address the first requirement, the study produced a suite of four competing candidate models, based on aggregations of the observations into different datasets. Specifically, Model 1 was based solely on variables for first signs and symptoms; Model 2 on blood biomarkers excluding PaO2/FiO2, and Model 3 and 4 using both sets of variables, including comorbidities. This experimental design enables a comparison of the relative predictive performance across categories of variables. Furthermore, the ranking of the variables by their predictive power makes it possible to achieve a parsimonious model by eliminating the least relevant variables, resulting in an effective yet parsimonious model.

To address the second and third requirements, the LightGBM suite of algorithms (Microsoft) [12] was used. These algorithms are based on well-known ensembles of Decision Trees, and are able to produce binary classification models (positive vs negative outcome) which tolerate missing data, and which support intelligible explanations on how the model achieves its predictions.

Meeting the final goal of minimizing FN (4) required that a specialized loss function be developed specifically for this task. This is a function that the algorithm must minimize in order to produce optimal predictions, and in this case it includes a tunable parameter to control the ratio of FP to FN.

For this binary classifier we used the PaO2/FiO2 ratio to derive a binary outcome for the learning task, where a positive outcome is defined as PaO2/FiO2 ≥ 150, and negative PaO2/FiO2 < 150. These are referred to as the positive and negative classes, respectively.

Following best practices, the dataset was divided into two parts: the training set (75% of the data—801 samples, of which 452 with PaO2/FiO2 < 150) and a complementary test set (the remaining 25% - 267 samples, of which 151 with PaO2/FiO2 < 150), which was not used in the learning phase. This separation was stratified according to the distribution of the outcome, in order to maintain a constant ratio between positive and negative classes in each of the subsets. The training set was used as input to the ML algorithm to train the model, while the test set is used to verify the predictive performance using standard metrics. The test set provided independent ground truth where each instance (a patient’s set of observations) was associated with one of the two possible outcomes. This test set was used to evaluate the predictive performance of the model, defined in terms of true positives (TP), true negatives (TN), false negatives (FN), false positives (FP).

Model performance was measured both on the AUROC and the sensitivity. The LightGBM algorithm not only allowed tuning of their hyper-parameters (these are the parameters that cannot be learnt by the algorithm and must be set manually) in order to maximize performance, but they also allowed more specific optimization targets than simply accuracy. For this application, clinical priority was followed to maximize the sensitivity, defined as TPTP+FN. Standard 5- and 10-fold cross validation was used to tune the model hyperparameters to achieve this goal.

To meet requirement (3) above, the learning framework provides explanations that go beyond the simple ranking of the variables. Specifically, the framework generates SHapley Additive exPlanations (SHAP) values quantifying the impact of each variable on the predicted outcome under different perspectives and both across the entire population and for individual patients [13].

Results

Out of a total of 295 patients, 198 with COVID-19 pneumonia were included. Clinical and laboratory characteristics are shown in Table 1. The vast majority of patients were males (69.2%) with a median age of 65 years. All patients showed relevant respiratory impairment as expressed by a median PaO2/FiO2: 262 mmHg (IQR: 150.0–316.0).

Table 1. Clinical and laboratory characteristics of the study population at the time of hospital admission.

Variable Mean or median or %
Population (%) 198 (100%)
Age, years, median (IRQ) 62.0 (54.0–73.0)
Sex, males (%) 151 (76.26%)
Signs and symptoms
Cough (%) 125 (63.33%)
Temperature, °C, median (IRQ) 37.0 (36.25–38.0)
Dyspnea (%) 135 (68.85%)
Diarrhea (%) 34 (17.19%)
Respiratory rate per minute, median (IRQ) 22.0 (18.75–28.0)
Biomarkers
Creatinine, mg/dl, median (IRQ) 0.92 (0.8–1.27)
D-Dimer, ng/ml, median (IRQ) 1120.0 (645.0–1767.5)
Hemoglobin, g/dl, median (IRQ) 13.85 (12.7–14.5)
Platelets, x109 per L, median (IRQ) 181.0 (154.5–258.5)
White Blood Cells, x1012 per L, median (IRQ) 6.92 (4.92–8.47)
PaO2, %, median (IRQ) 65.2 (56.9–76.6)
PaO2/FiO2, mmHg, median (IRQ) 256.0 (149.75–301.5)
Oxygen saturation, %, median (IRQ) 94.0 (91.65–95.75)
Creatinine kinase, U/L, median (IRQ) 107.0 (60.25–261.75)
Lactate dehydrogenase, U/L, median (IRQ) 562.0 (458.0–663.0)
C-reactive protein, mg/dl, median (IRQ) 7.35 (4.22–17.65)

Abbreviations: IQR–interquartile range; FiO2 –fraction of inspired oxygen; PaO2 –partial arterial pressure of oxygen.

To recollect, all models are binary (yes/no) classifiers of risk of developing respiratory failure, measured using a quantitative criterion (PaO2/FiO2 < 150 mmHg) and assessed using AUC and sensitivity. Also, a specific loss function was developed to privilege models that minimize the number of false negatives (FN).

Table 2 describes the four models that were used to train and test the ML tool with the following features:

Table 2. Describes the four models that were used to train and test the ML tool.

Table legend AUROC (%)
Actual Good Outcome Actual Good Outcome Training model: %
Actual Good Outcome TN FP
Test model: %
Actual Bad Outcome FN TP
Model 1 –Only signs and symptoms, 31 variables Model 1
Actual Good Outcome Actual Good Outcome Training model: 89%
Actual Good Outcome 74—TN 42 FP
Test model: 69%
Actual Bad Outcome 56 FN 95 TP
Model 2 –Laboratory biomarkers, 39 variables Model 2
Predicted Good Outcome Predicted Bad Outcome Training model: 97%
Actual Good Outcome 89 27
Test model: 83%
Actual Bad Outcome 39 112
Model 3 –Extended mixed model, 91 variables Model 3
Predicted Good Outcome Predicted Bad Outcome Training model: 99%
Actual Good Outcome 95 21
Test model: 85%
Actual Bad Outcome 40 111
Model 4 –Boosted mixed model, 20 variables Model 4
Predicted Good Outcome Predicted Bad Outcome Training model: 98%
Actual Good Outcome 87 29
Test model: 84%
Actual Bad Outcome 40 111
  • Model 1: “signs and symptoms” that included 31 variables

  • Model 2: “laboratory biomarkers” that included 39 variables.

  • Model 3: “extended mixed model” that included 91 variables.

  • Model 4: “boosted mixed model” that includes 20 variables.

The latter was obtained from Model 3, achieving the same performance, and equal number of FN, with only 20 variables out of 91 of Model 3. To achieve this result, the variables of Model 3 were ranked according to their mean SHAP values and a backward variable selection approach was adopted, by which each variable was excluded in turn, starting from the lowest-ranked variables, and the loss in predictive performance was measured each time. This procedure identified Model 4 as the one that minimized the number of variables (20 out of the original 91) without worsening the FN rate, while achieving the best performance (AUROC = 0.84). S1 Fig specifies the description of the proportion of available data for each of the 20 variables.

For each model, Table 2 reports the Area Under the Curve (AUC) for both training and test sets. We indicate the number of TP, TN, FN, FP. Variables included in each model are listed in S1 File. Figs 1 and 2 show the top 20 ranking variables used to train Model 4. X axes show the average impact of model output magnitude, expressed by SHAP values.

Fig 1. Shows top 20 ranking variables used to train Model 4.

Fig 1

X axes show the average impact of model output magnitude, expressed by SHAP values.

Fig 2. The individual SHAP values for the 20 top variables.

Fig 2

Values of each variable may have a positive or negative impact depending on their SHAP value, for instance high values of dyspnea in red contribute strongly to the positive class (negative patient outcome), while low values in blue contribute strongly to the negative class (positive outcome).

Case presentation to support customization of ML into clinical practice

In addition to the theoretical performance, anecdotal validation was also provided in a real-life case setting. We applied our model to the clinical course of a 55-year-old male who was admitted for high fever and shortness of breath due to COVID-19 pneumonia. Antiviral therapy with darunavir/cobicistat was started in addition to hydroxychloroquine. He was discharged the following day, in the absence of respiratory failure as assessed by PaO2/FiO2 = 420 mmHg. Four days later, he was readmitted to hospital with high fever (39°C), diarrhea and onset of mild respiratory failure (PaO2/FiO2 = 230 mmHg). Inflammatory biomarkers were high (CRP 18 mg/dl) with elevated neutrophils. In the following 24 hours, the patient experienced a clinically unpredictable dramatic worsening of his clinical condition due to the onset of severe respiratory distress despite adequate oxygen supply (PaO2/FiO2 = 88 mmHg, respiratory rate higher than 35 breaths per minute). He was then transferred to the Intensive Care Unit (ICU) where non-invasive mechanical ventilation with helmet in pressure support mode was initiated. After 8 days of assisted spontaneous breathing, he was weaned from NIV and discharged the following day without oxygen supply (Fig 3).

Fig 3. Implementation of machine learning model in the case of 55-year old patient who was admitted and discharged the following day.

Fig 3

On day 4, the patient was re-admitted with mild respiratory insufficiency that had a 87.7% probability to experience a respiratory crush in the following 48 hours.

As shown in the case scenario presented above, the model was retrospectively applied in order to explore the prediction of “respiratory crush” in support of clinical judgment. From the physician’s point of view, the first discharge was motivated by the stable clinical conditions. However, our model showed a 36.6% probability of worsening of the respiratory function in the following 48 hours, meeting the criteria for mechanical ventilation with pressure support in the next two days. Moreover, the model was able to predict at the time of the second admission, the respiratory function decline that our patient actually experienced 24 hours later. The model at day 14 predicted a 47.4% risk of new worsening, but this should have been integrated with clinical data suggested by patient’s perception of improvement and rapid increase of blood gas exchange. A development of our support model at the time into our clinical practice would have provided support to clinical judgment, suggesting against the first discharge, and furthermore, recommending continuous monitoring once the patient was readmitted in order to possibly avoid ICU with urgent treatment.

Discussion

We have created a statistical learning model to assist clinicians in forecasting patients with COVID-19 who develop respiratory failure requiring mechanical ventilation. The model provide a reliable 48 hours prediction of moderate to severe respiratory failure, with an accuracy of 84% that minimizes the FN rate.

The level of performance of our model is in line with other ML tools used in different areas of medicine [14, 15] and it is very useful in COVID-19 clinical context where disease progression remains unpredictable both in the early virologic and in the late inflammatory phase.

It must be acknowledged that risk probability generated from the algorithm at each time point is not a measure of overall performance of the model. Clinicians should not interpret the punctual probability score as a diagnosis but rather to asses the trend measure, integrating the data in the context of clinical judgment.

We chose to have a short-term outcome to support clinicians at hospital admission and discharge. Given the rapid and dynamic clinical changes affecting COVID-19 patients, this time frame should be considered crucial for the initiation of therapies aimed at avoiding ICU admission and mechanical ventilation. In the future we may be able to develop similar models to also support clinicians to better interpret patient’s clinical improvement after they are discharged from hospital.

The construction of different models followed a clinically oriented variables choice. The first model based on 31 variables obtained from signs and symptoms returned a suboptimal prediction accuracy. Adding biomarkers including respiratory variables significantly increased the forecasting capacity of the model. The best performance was obtained in the boosted mixed model, which however still requires about 20 variables. From a physician’s perspective, a cluster of 20 variables may be difficult to manage in routine clinical practice. What our approach offers in support to the decision-making process is a simple interpretation of the predictions.

Machine learning approach is at the top of the list of the research priorities proposed by the Horizon 2020 program (H2020) [16]. This study may contribute to develop new analytics to improve care at high technology readiness levels.

Moderate to severe respiratory failure was chosen as an outcome being the most relevant time point in the natural history of severe COVID-19 pneumonia. At a clinical level, it represents the so-called “respiratory crush” which marks acute lung injury and leads to mechanical ventilation in ICU. At a public health level, this machine learning model might be helpful in optimizing scarce resources like ventilators and ICU beds.

A few clinical risk scores have been developed and validated to predict the occurrence of critical illness in hospitalized patients with COVID-19. These scores used at time of admission either the neutrophil/lymphocyte ratio or 10 clinical variables including radiological findings to predict critical illness using a traditional statistical approach to generate a prediction algorithm [17, 18].

In a similar experience from China, ML was used to predict mortality in patients with COVID-19, using three biomarkers only [7].

Recent data suggests that COVID-19 does not affect only respiratory system, but also other organs, such as liver, kidneys, gut, heart and central nervous system [19].

Given the multisystemic nature of this disease, limited number of parameters may not be sufficient to predict worsening in these patients.

Not surprisingly, this hard endpoint can be predicted with a very limited number of biomarkers, reducing the clinical parameters to be monitored. However, clinical worsening seems to be more challenging to forecast. An intermediate dynamic event with multiple biomarkers appears to be more difficult to predict than a final static event, such as mortality, with a small number of variables.

This science data faced several methodological challenges. Features which fed the model were chosen based both on the Shapley Values approach and on clinicians' suggestion, in a hybrid approach. This allowed to take advantage of both aspects: on one side the clinical experience of physicians who selected variables and outcomes using a knowledge-based approach, and on the other side, the probabilistic nature of a data-driven framework.

Microsoft LightGBM framework was chosen in particular to support missing data deriving from a clinical setting where it was not practical to collect all observations at each data point. Clinicians appreciated the “Glass box” opportunity, which showed the top variables, trusting a model in which pathophysiological interpretation could still be plausible. With regards to the 20 variables selected in the hybrid models, some can be clustered within the hypoxic damage (dyspnea, HCO3-, pH, reparatory and heart rate) other in relation to inflammation (C-reactive protein, D-dimer, platelets, red blood cells, lymphocytes), other in relation to organ damage (lactate dehydrogenase, creatinine-kinase). Medical Decision Support Systems must provide transparency to explain how the predictive model behaves. In this perspective, an interpretation approach is necessary, both to have better understanding of the patient's health status and to better identify dangerous biases that the model could have learnt.

The approach in our study is substantially different from traditional models based on logistic regression [20]. Risk scores derived from logistic regression mixed effect models are knowledge-driven approaches where a score is assigned by an expert to each of the limited number of selected variables [21]. In contrast, our predictive model is data driven. It is based on Decision Trees, in which the relative effect of all available variables is considered with their raw value instead of an arbitrary summarization.

This approach is particularly appropriate in COVID-19 context as it aligns well with principles of personalized medicine. It allows to identify what are the most relevant biomarkers for each individual patient which in a single patient can lead to a favorable or un-favorable progression, as shown by the Shapley Values application in Fig 3. This confirms the expected relative importance (ranking) of the variables used as features and the roles of those features in the model. This is a further proof that the model is robust as it relies on the values of variables deemed relevant in this context without actually knowing their semantics.

Our study has several limitations. Firstly, it does not take account of radiological findings on X-rays or CT, as these data were not collected consistently during hospital stay. Nevertheless, a recent radiological study demonstrated only a slight increase of prediction of respiratory failure adding thoracic CT to clinical data [22].

Secondly, we chose as outcome a cut-off of PaO2/FiO2 which represents one of the most important criteria for invasive mechanical ventilation, but not mechanical ventilation itself. Also, the model is oblivious to a whole panel of interleukins values, as they were not collected on daily basis regardless of their potential involvement in the development of severe respiratory failure. Lastly, the model will need to evolve with the growth of the dataset, providing a more accurate cut-off risk value.

In conclusion, this study developed a machine learning algorithm aimed to assist clinicians in dealing with COVID-19 health emergency. It is proving useful in predicting severe respiratory failure requiring mechanical ventilation in the following 48 hours, allowing to anticipate urgent events potentially improving management of critically ill patients.

Supporting information

S1 Fig

(PDF)

S1 File

(DOCX)

Acknowledgments

We would like to thank to Office of clinical protocols and data management in Modena: Barbara Beghetto, Giulia Nardini, Enrica Roncaglia; Office of information and communication technologies of Policlinico di Modena: Rossella Fogliani, Grazia Righini, Mario Lugli.

Data Availability

All relevant data are within the manuscript and its Supporting Information files.

Funding Statement

The author(s) received no specific funding for this work.

References

  • 1.Zhou F, Yu T, Du R, et al. Clinical course and risk factors for mortality of adult inpatients with COVID-19 in Wuhan, China: a retrospective cohort study. Lancet. 2020; 6736(20):1–9. 10.1016/S0140-6736(20)30566-3 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 2.Nicastri E, D’Abramo A, Faggioni G, et al. Coronavirus disease (COVID-19) in a paucisymptomatic patient: epidemiological and clinical challenge in settings with limited community transmission, Italy, February 2020. Euro Surveill Bull Eur sur les Mal Transm = Eur Commun Dis Bull. Sweden; 2020; 25(11). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 3.Guan W-J, Ni Z-Y, Hu Y, et al. Clinical Characteristics of Coronavirus Disease 2019 in China. N Engl J Med. United States; 2020. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 4.Meng L, Qiu H, Wan L, et al. Intubation and Ventilation amid the COVID-19 Outbreak: Wuhan’s Experience. Anesthesiology. United States; 2020. 10.1097/ALN.0000000000003296 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 5.Shortliffe EH, Sepulveda MJ. Clinical Decision Support in the Era of Artificial Intelligence. JAMA. United States; 2018; 320(21):2199–2200. 10.1001/jama.2018.17163 [DOI] [PubMed] [Google Scholar]
  • 6.Hamet P, Tremblay J. Artificial intelligence in medicine. Metabolism. United States; 2017; 69S:S36–S40. 10.1016/j.metabol.2017.01.011 [DOI] [PubMed] [Google Scholar]
  • 7.Yan L, Zhang H-T, Goncalves J, et al. A machine learning-based model for survival prediction in patients with severe COVID-19 infection. medRxiv [Internet]. 2020;: 2020.02.27.20028027. Available from: http://medrxiv.org/content/early/2020/03/17/2020.02.27.20028027.abstract [Google Scholar]
  • 8.Tárnok A. Machine Learning, COVID-19 (2019-nCoV), and multi-OMICS. 2020; 19:215–216. 10.1002/cyto.a.23990 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 9.Wynants L, Calster B Van, Bonten MMJ, et al. Prediction models for diagnosis and prognosis of covid-19 infection: systematic review and critical appraisal. BMJ. England; 2020; 369:m1328 10.1136/bmj.m1328 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 10.National Society of Infectious Diseases’ (SIMIT) Guidelines for the managment of COVID-19 infection. Available at: http://www.simit.org/medias/1569-covid19-vademecum-13-03-202.pdf Last access April 5th 2020. No Title.
  • 11.Cao B, Wang Y, Wen D, et al. A Trial of Lopinavir-Ritonavir in Adults Hospitalized with Severe Covid-19. N Engl J Med [Internet]. 2020;: 1–13. Available from: http://www.ncbi.nlm.nih.gov/pubmed/32187464 10.1056/NEJMoa2001282 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 12.Ke G, Meng Q, Finley T, et al. LightGBM: A Highly Efficient Gradient Boosting Decision Tree. In: Guyon I, Luxburg U V, Bengio S, et al., editors. Adv Neural Inf Process Syst 30 [Internet]. Curran Associates, Inc.; 2017. p. 3146–3154. Available from: http://papers.nips.cc/paper/6907-lightgbm-a-highly-efficient-gradient-boosting-decision-tree.pdf
  • 13.Lundberg SM, Lee S-I. A Unified Approach to Interpreting Model Predictions. In: Guyon I, Luxburg U V, Bengio S, et al., editors. Adv Neural Inf Process Syst 30 [Internet]. Curran Associates, Inc.; 2017. p. 4765–4774. Available from: http://papers.nips.cc/paper/7062-a-unified-approach-to-interpreting-model-predictions.pdf
  • 14.Yasaka K, Abe O. Deep learning and artificial intelligence in radiology: Current applications and future directions. PLoS Med. 2018. p. e1002707 10.1371/journal.pmed.1002707 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 15.Kumamaru KK, Fujimoto S, Otsuka Y, et al. Diagnostic accuracy of 3D deep-learning-based fully automated estimation of patient-level minimum fractional flow reserve from coronary computed tomography angiography. Eur Hear J—Cardiovasc Imaging [Internet]. 2019; 21(4):437–445. Available from: 10.1093/ehjci/jez160 [DOI] [PubMed] [Google Scholar]
  • 16.European Research Area corona platform; Available at: https://ec.europa.eu/info/funding-tenders/opportunities/portal/screen/opportunities/topic-details/sc1-phe-coronavirus-2020-2b. Last access 29 May 2020.
  • 17.Liang W, Liang H, Ou L, et al. Development and Validation of a Clinical Risk Score to Predict the Occurrence of Critical Illness in Hospitalized Patients With COVID-19. JAMA Intern Med. 2020. 10.1001/jamainternmed.2020.2033 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 18.Liu J, Liu Y, Xiang P, et al. Neutrophil-to-lymphocyte ratio predicts critical illness patients with 2019 coronavirus disease in the early stage. J Transl Med. 2020; 18(1):206 10.1186/s12967-020-02374-0 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 19.Gupta A, Madhavan M V, Sehgal K, et al. Extrapulmonary manifestations of COVID-19. Nat Med [Internet]. 2020; 26(7):1017–1032. Available from: 10.1038/s41591-020-0968-3 [DOI] [PubMed] [Google Scholar]
  • 20.Desai RJ, Wang S V, Vaduganathan M, Evers T, Schneeweiss S. Comparison of Machine Learning Methods With Traditional Models for Use of Administrative Claims With Electronic Medical Records to Predict Heart Failure Outcomes. JAMA Netw open. United States; 2020; 3(1):e1918962. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 21.Mohammed MA, Moles RJ, Hilmer SN, Kouladjian O’Donnel L, Chen TF. Development and validation of an instrument for measuring the burden of medicine on functioning and well-being: The Medication-Related Burden Quality of Life (MRB-QoL) tool. BMJ Open. 2018; 8(1):1–11. 10.1136/bmjopen-2017-018880 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 22.Colombi D, Bodini FC, Petrini M, et al. Well-aerated Lung on Admitting Chest CT to Predict Adverse Outcome in COVID-19 Pneumonia. Radiology. 2020; 296(2):E86–E96. 10.1148/radiol.2020201433 [DOI] [PMC free article] [PubMed] [Google Scholar]

Decision Letter 0

Paola Faverio

3 Aug 2020

PONE-D-20-16454

Machine learning in predicting respiratory failure in patients with COVID-19 pneumonia - challenges, strengths, and opportunities in a global health emergency

PLOS ONE

Dear Dr. Guaraldi,

Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process.

I apologise for the long review process due, as you will understand, to the historical moment and its compelling commitments.

Based on the comments from the reviewers and my personal revision I suggest minor revisions to be made.

Please submit your revised manuscript by August 22nd. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at plosone@plos.org. When you're ready to submit your revision, log on to https://www.editorialmanager.com/pone/ and select the 'Submissions Needing Revision' folder to locate your manuscript file.

Please include the following items when submitting your revised manuscript:

  • A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.

  • A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.

  • An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'.

If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter.

If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: http://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory-protocols

We look forward to receiving your revised manuscript.

Kind regards,

Paola Faverio

Academic Editor

PLOS ONE

Additional Editor Comments:

Can the authors provide more information on the validation of the machine learning method used in this study?

Journal requirements:

When submitting your revision, we need you to address these additional requirements.

1. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at

https://journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_main_body.pdf and

https://journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatting_sample_title_authors_affiliations.pdf

2. Please clarify in your Methods section whether your study was prospective or retrospective, and whether any intervention was applied. Please also clarify in your Methods section the name of the ethics committee and approval number. Please also provide details of participant consent, and whether this was written or verbal.

3.Please clarify in your Data availability statement how other researchers may obtain the data used in the study. Please also clarify whether the model code has been or will be made available to other researchers.

4. *Please explain the rationale for the development of your model in light of recent research in this area, clearly indicating which problem with existing models you are addressing.

*Please clearly report at the beginning of your methods or results section which were the key performance measures used to establish the validity and utility of your model. Please also report clearly which statistical analysis was used to establish robustness of performance measures.

*Please note that PLOS ONE requires that experiments, statistics, and other analyses must be performed to a high technical standard and described in sufficient detail to allow for reproducibility of the study (http://journals.plos.org/plosone/s/criteria-for-publication#loc-3). To demonstrate the performance of the model, we would expect comparisons to be drawn between existing state-of-the-art methods.

5. Please upload a copy of Supporting Information Table 1 which you refer to in your text on page 8.

[Note: HTML markup is below. Please do not edit.]

Reviewers' comments:

Comments to the Author

1. Is the manuscript technically sound, and do the data support the conclusions?

The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented.

Reviewer #1: Yes

Reviewer #2: Yes

**********

2. Has the statistical analysis been performed appropriately and rigorously?

Reviewer #1: Yes

Reviewer #2: Yes

**********

3. Have the authors made all data underlying the findings in their manuscript fully available?

The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified.

Reviewer #1: Yes

Reviewer #2: No

**********

4. Is the manuscript presented in an intelligible fashion and written in standard English?

PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here.

Reviewer #1: Yes

Reviewer #2: Yes

**********

5. Review Comments to the Author

Reviewer #1: This is an interesting observational study about the use the prediction of respiratory failure in patients with COVID-19 pneumonia.

In my personal opinion, the principal limitation of the study is the lack of important data such as radiological included in the prediction that will be very important in this specific population. However, the study presents an important data and easy model to predict respiratory failure in COVID.19 patients with pneumonia. Finally, the inclusion and exclusion criteria are not totally clear in my opinion. This is an interesting and new issue.

I have minor comments:

1.- Please unify the use of the term COVID-19 throughout the manuscript

2.- Could the inclusion and exclusion criteria be better explained?

3.- Did you have the PaO2 / FiO2 data for all the patients? you had some specific protocol to have these data from all patients consecutively admitted to the emergency department or the authors included only patients who needed ICU admission

Reviewer #2: This manuscript shows the clinical utility of learning algorithms in the prognostication of COVID19 pneumonia. The Authors offer a prediction model for COVID19 patients, however I felt the main achievement of this project is to demonstrate that machine learning can be readily helpful in clinical practice for the chest physician. In the respiratory field, machine learning techniques have mainly been researched in lung cancer and chest imaging; it is probably time to explore its potential somewhere else.

There are some minor issues/suggestions:

1) In the abstract: “pao2/fio2 ratio < 150 in at least one of two consecutive ABG in the following 48h” - this should be included in the Methods section of the full text (there is something similar in the “study design” paragraph, but this is far clearer)

2) “Missing data” is mentioned several times: please provide a description of the missing data (a short summary or a graph or n. of variables with missing > 15% or 25%). It should be fine to add it as supplementary text, if allowed.

3) I was not able to find the number of patients with positive/negative outcome (paO2/FiO2 >/< 150) in both subsets: this should be provided

4) No information regarding past medical history/comorbidities: this can have an important impact on outcomes such as pao2/Fio2 status (e.g. chronic respiratory conditions). Please briefly mention this issue.

5) Discussion section, 2nd paragraph starting with “Our model is (1)...”: this is already in the methods section (remove it, or just refer to the methods)

6) Discussion, 2nd page “It might allow to optimize….”: I think I understood your point, but what is the subject? (Efforts to cut the progression to “respiratory crush” might allow to optimize…??)

7) Results section, last paragraph: “the model at day 14 predicted a 47% risk”, please make clear this was an error (something like “erroneously predicted 47%...”)

8) Results section, last paragraph “but this should have been integrated with clinical data….”: please remove this sentence (or move it to the discussion section) - results should not be commented

9) Results, last sentence starting with “A deployment of our support model….”: please just correct the english typo

10) Discussion session: looking at the baseline tables, data seem to be skewed toward less severe patients (see median values of LDH/D-Dimer/CRP/pao2-fio2 ratio); this may have an impact for the reproducibility of the results and model performance as well

11) Discussion session, sentence starting with “Not surprisingly this hard endpoint….”: I understand the Authors’ view, but not necessarily mortality is easier to predict than “softer” disease progression - It depends when we start following the patient for example, and mortality is of course “disease progression”. Please, rewrite the sentence with less emphasys; if possible try to add an alternative explanation for the higher number of variables needed compared to [7]

12) Discussion section, 2nd page: “were chosen based both on a statistical exploratory data analysis and on clinicians’ suggestion” - this should be stated in the methods section as well, adding some details about the type of “data analysis” used to select variable for the ML model

13) COVID19 outcome prediction models with fewer elements and similar diagnostic accuracy have been developed using a more “traditional” approach (LASSO, logistic…) [e.g. Wenhua Liang et al, JAMA 2020, PMID 32396163; Jingyuan Liu et al, J Transl Med 2020, PMID 32434518]. In order to give a broader perspective to the interested reader, please mention that in the discussion section along with your interpretation.

14) Figure 1 & 2: there is a typo (“dispnea”)

**********

6. PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files.

If you choose “no”, your identity will remain anonymous but your review may still be made public.

Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy.

Reviewer #1: No

Reviewer #2: No

[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.]

While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, https://pacev2.apexcovantage.com/. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at figures@plos.org. Please note that Supporting Information files do not need this step.

PLoS One. 2020 Nov 12;15(11):e0239172. doi: 10.1371/journal.pone.0239172.r002

Author response to Decision Letter 0


21 Aug 2020

Modena, 14 August 2020

Dear Editor,

We are very grateful for your constructive comments and suggestions to our paper entitled: “Machine learning in predicting respiratory failure in patients with COVID-19 pneumonia - challenges, strengths, and opportunities in a global health emergency”.

We here provide a point-by-point reply to the comments and we have incorporated the related changes in the manuscript. We thank the reviewers for their thoughtful insights which helped to significantly improve the manuscript.

PONE-D-20-16454

Machine learning in predicting respiratory failure in patients with COVID-19 pneumonia - challenges, strengths, and opportunities in a global health emergency

PLOS ONE

Dear Dr. Guaraldi,

Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process.

I apologise for the long review process due, as you will understand, to the historical moment and its compelling commitments.

Based on the comments from the reviewers and my personal revision I suggest minor revisions to be made.

Please submit your revised manuscript by August 22nd. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at plosone@plos.org. When you're ready to submit your revision, log on to https://www.editorialmanager.com/pone/ and select the 'Submissions Needing Revision' folder to locate your manuscript file.

Please include the following items when submitting your revised manuscript:

• A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.

• A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.

• An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'.

If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter.

If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.ioassigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: http://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory-protocols

We look forward to receiving your revised manuscript.

Kind regards,

Paola Faverio

Academic Editor

PLOS ONE

Additional Editor Comments:

Can the authors provide more information on the validation of the machine learning method used in this study?

Authors’ response:

The following paragraphs refer to the validation of our model:

“Model performance was measured both on the AUROC and the sensitivity. The LightGBM algorithm not only allowed tuning of their hyper-parameters (these are the parameters that cannot be learnt by the algorithm and must be set manually) in order to maximize performance, but they also allowed more specific optimization targets than simply accuracy. For this application, clinical priority was followed to maximize the sensitivity, defined as TP/(TP+FN). Standard 5- and 10-fold cross validation was used to tune the model hyperparameters to achieve this goal.

To meet requirement (3) above, the learning framework provides explanations that go beyond the simple ranking of the variables. Specifically, the framework generates SHapley Additive exPlanations (SHAP) values quantifying the impact of each variable on the predicted outcome under different perspectives and both across the entire population and for individual patients [13].”

Journal requirements:

When submitting your revision, we need you to address these additional requirements.

1. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at

https://journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_main_body.pdf and

https://journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatting_sample_title_authors_affiliations.pdf

2. Please clarify in your Methods section whether your study was prospective or retrospective, and whether any intervention was applied. Please also clarify in your Methods section the name of the ethics committee and approval number. Please also provide details of participant consent, and whether this was written or verbal.

Authors’ response:

According to Editor’s suggestion, this has been added in the methods.

3.Please clarify in your Data availability statement how other researchers may obtain the data used in the study. Please also clarify whether the model code has been or will be made available to other researchers.

Authors’ response:

Data can be made available subject to a data disclosure agreement to be arranged with the Policlinico Universitario di Modena e Reggio Emilia.

4. *Please explain the rationale for the development of your model in light of recent research in this area, clearly indicating which problem with existing models you are addressing.

Authors’ response:

We have added an explanation for the rationale to the background section. This starts by highlighting the findings from the systematic review [9].

Essentially, the review finds that some risk prediction models do exist that attempt to predict risk of intensive care unit admission, ventilation, intubation. The review concludes that most of these studies have shortcomings (high bias, poor reporting) that make them unsuitable for clinical decision-making. In contrast, the models presented in this work are explainable, meaning that they provide an easily understandable grounding for choice of predictors and their relative importance on individual outcomes.

*Please clearly report at the beginning of your methods or results section which were the key performance measures used to establish the validity and utility of your model. Please also report clearly which statistical analysis was used to establish robustness of performance measures.

Authors’ response:

The models are binary (yes/no) classifiers of risk of developing respiratory failure, measured using a quantitative criterion (PaO2/FiO2 < 150 mmHg) and assessed using AUC and sensitivity. Also, a specific loss function was developed to privilege models that minimize the number of false negatives (FN).

This phrasing has been added at the beginning of the Results section.

*Please note that PLOS ONE requires that experiments, statistics, and other analyses must be performed to a high technical standard and described in sufficient detail to allow for reproducibility of the study (http://journals.plos.org/plosone/s/criteria-for-publication#loc-3). To demonstrate the performance of the model, we would expect comparisons to be drawn between existing state-of-the-art methods.

Authors’ response:

Statistics, and other analyses were performed in a high technical standard described in the methods.

5. Please upload a copy of Supporting Information Table 1 which you refer to in your text on page 8.

Authors’ response:

Supplementary table 1 is now added listing in alphabetical order the 91 variables used to build the model.

[Note: HTML markup is below. Please do not edit.]

Reviewers' comments:

Comments to the Author

1. Is the manuscript technically sound, and do the data support the conclusions?

The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented.

Reviewer #1: Yes

Reviewer #2: Yes

2. Has the statistical analysis been performed appropriately and rigorously?

Reviewer #1: Yes

Reviewer #2: Yes

3. Have the authors made all data underlying the findings in their manuscript fully available?

The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified.

Reviewer #1: Yes

Reviewer #2: No

4. Is the manuscript presented in an intelligible fashion and written in standard English?

PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here.

Reviewer #1: Yes

Reviewer #2: Yes

5. Review Comments to the Author

Reviewer #1: This is an interesting observational study about the use the prediction of respiratory failure in patients with COVID-19 pneumonia.

In my personal opinion, the principal limitation of the study is the lack of important data such as radiological included in the prediction that will be very important in this specific population.

Authors’ response:

We agree with this comment and in the limitation of the study it is written” Our study has several limitations. Firstly, it does not take account of radiological findings on X-rays or CT,”. Nevertheless some radiological studies underline that the contribution of radiological findings in the prediction of subsequent Acute Respiratory Distress Syndrome in minimal. In the case of Colombi study, CT findings mimimally increased AUC prediction from 0,83 to 0,85 in addition to clinical data. This sentence is added:

“Our study has several limitations. Firstly, it does not take account of radiological findings on X-rays or CT, as these data were not collected consistently during hospital stay. Nevertheless, a recent radiological study demonstrated only a slight increase of prediction of respiratory failure adding thoracic CT to clinical data.”

However, the study presents an important data and easy model to predict respiratory failure in COVID.19 patients with pneumonia. Finally, the inclusion and exclusion criteria are not totally clear in my opinion.

Authors’ response:

The following sentence has been added in the methods:

“Hospitalized patients with COVID-19 pneumonia were included if they had at least two arterial blood gas analyses measurements in the following 48 hours.”

This is an interesting and new issue.

I have minor comments:

1.- Please unify the use of the term COVID-19 throughout the manuscript

Authors’ response:

This has been corrected.

2.- Could the inclusion and exclusion criteria be better explained?

Authors’ response:

Please, see below.

3.- Did you have the PaO2 / FiO2 data for all the patients? you had some specific protocol to have these data from all patients consecutively admitted to the emergency department or the authors included only patients who needed ICU admission

Authors’ response:

Yes, this was specified in the inclusion criteria.

Reviewer #2: This manuscript shows the clinical utility of learning algorithms in the prognostication of COVID19 pneumonia. The Authors offer a prediction model for COVID19 patients, however I felt the main achievement of this project is to demonstrate that machine learning can be readily helpful in clinical practice for the chest physician. In the respiratory field, machine learning techniques have mainly been researched in lung cancer and chest imaging; it is probably time to explore its potential somewhere else.

There are some minor issues/suggestions:

1) In the abstract: “pao2/fio2 ratio < 150 in at least one of two consecutive ABG in the following 48h” - this should be included in the Methods section of the full text (there is something similar in the “study design” paragraph, but this is far clearer)

Authors’ response:

This sentence has been improved according to reviewer suggestion:

The study outcome was the onset of moderate to severe respiratory failure defined as PaO2/FiO2 ratio < 150 mmHg (≤ 13.3 kPa) in at least one of two consecutive arterial blood gas analyses in the following 48 hours.

2) “Missing data” is mentioned several times: please provide a description of the missing data (a short summary or a graph or n. of variables with missing > 15% or 25%). It should be fine to add it as supplementary text, if allowed.

Authors’ response:

This sentence and table have been added in the text:

Supplementary Figure 2 specifies the description of the proportion of available data for each of the 20 variables.

3) I was not able to find the number of patients with positive/negative outcome (paO2/FiO2 >/< 150) in both subsets: this should be provided

Authors’ response:

Study outcome was identified from PaO2/FiO2 data and individual patients could contribute to both positive and negative outcomes. The same patient could contribute to generate the definition of respiratory failure or the lack of this condition, therefor it may be misleading to specify the number of patients in non-mutually exclusive groups. We preferred to specify in the text the number of observations rather than the number of patients. It was written: “In detail, 603 observations contributed to the definition of respiratory failure (PaO2/FiO2 < 150 mmHg) and 465 did not meet this definition”.

4) No information regarding past medical history/comorbidities: this can have an important impact on outcomes such as pao2/Fio2 status (e.g. chronic respiratory conditions). Please briefly mention this issue.

Authors’ response:

History of comorbidities was available in a subset of 119 patients and was included in the training and test set of machine learning prediction, as specified in methods.

5) Discussion section, 2nd paragraph starting with “Our model is (1) ...”: this is already in the methods section (remove it, or just refer to the methods)

Authors’ response:

This sentence has been removed.

6) Discussion, 2nd page “It might allow to optimize….”: I think I understood your point, but what is the subject? (Efforts to cut the progression to “respiratory crush” might allow to optimize…??)

Authors’ response:

This sentence has been corrected as following:

“At a public health level, this machine learning model might be helpful in optimizing scarce resources like ventilators and ICU beds.”

7) Results section, last paragraph: “the model at day 14 predicted a 47% risk”, please make clear this was an error (something like “erroneously predicted 47%...”)

Authors’ response:

47% risk is the risk probability at that time point and not a measure of performance of the model. Each time point probability, as mentioned thereafter should have been integrated with clinical data. We state “but” to stress the need to integrate prediction and clinical data.

The following sentence was added:

“It must be acknowledged that risk probability generated from the algorithm at each time point is not a measure of overall performance of the model. Clinicians should not interpret the punctual probability score as a diagnosis but rather to assess the trend measure, integrating the data in the context of clinical judgment”.

8) Results section, last paragraph “but this should have been integrated with clinical data….”: please remove this sentence (or move it to the discussion section) - results should not be commented

Authors’ response:

This result is a case example which explain probability risk and overall performance of the model and do not fit elsewhere the discussion section. We respectfully suggest to leave this small comment in this section.

As mentioned above this sentence was added in the discussion:

“It must be acknowledged that risk probability generated from the algorithm at each time point is not a measure of overall performance of the model. Clinicians should not interpret the punctual probability score as a diagnosis but rather assessing the trend measure integrating the data in the context of clinical judgment.”

9) Results, last sentence starting with “A deployment of our support model….”: please just correct the english typo

Authors’ response:

This has been corrected.

10) Discussion session: looking at the baseline tables, data seem to be skewed toward less severe patients (see median values of LDH/D-Dimer/CRP/pao2-fio2 ratio); this may have an impact for the reproducibility of the results and model performance as well

Authors’ response:

A certain amount of skew in a multi-dimensional space of features (the variables in Table I) is inevitable. Regarding model performance, skew may affect generalization error, that is it may result in model overfitting. We control overfitting according to standard Machine Learning practices, namely by (1) deploying ensemble methods that are known to be robust to data skew, i.e., LightGBM / XGBoost in this instance, (2) using K-fold cross-validation to assess model performance, and (3) computing model performance on an independent test set that was not used for training.

11) Discussion session, sentence starting with “Not surprisingly this hard endpoint….”: I understand the Authors’ view, but not necessarily mortality is easier to predict than “softer” disease progression - It depends when we start following the patient for example, and mortality is of course “disease progression”. Please, rewrite the sentence with less emphasis; if possible try to add an alternative explanation for the higher number of variables needed compared to [7]

Authors’ response:

The sentence has been introduced as following:

“Recent data suggests that COVID-19 does not affect only respiratory system, but also other organs, such as liver, kidneys, gut, heart and central nervous system [19]. Given the multisystemic nature of this disease, limited number of parameters may not be sufficient to predict worsening in these patients.

Not surprisingly, this hard endpoint can be predicted with a very limited number of biomarkers, reducing the clinical parameters to be monitored. However, clinical worsening seems to be more challenging to forecast. An intermediate dynamic event with multiple biomarkers appears to be more difficult to predict than a final static event, such as mortality, with a small number of variables.”

12) Discussion section, 2nd page: “were chosen based both on a statistical exploratory data analysis and on clinicians’ suggestion” - this should be stated in the methods section as well, adding some details about the type of “data analysis” used to select variable for the ML model

Authors’ response:

This sentence has been modified in the discussion.

This science data faced several methodological challenges. Features which fed the model were chosen based both on the Shapley Values approach and on clinicians' suggestion, in a hybrid approach.

13) COVID19 outcome prediction models with fewer elements and similar diagnostic accuracy have been developed using a more “traditional” approach (LASSO, logistic…) [e.g. Wenhua Liang et al, JAMA 2020, PMID 32396163; Jingyuan Liu et al, J Transl Med 2020, PMID 32434518]. In order to give a broader perspective to the interested reader, please mention that in the discussion section along with your interpretation.

Authors’ response:

We thank the reviewer for the valuable references that have been introduced in the discussion:

“A few clinical risk scores have been developed and validated to predict the occurrence of critical illness in hospitalized patients with COVID-19. These scores used at time of admission either the neutrophil/lymphocyte ratio or 10 clinical variables including radiological findings to predict critical illness using a traditional statistical approach to generate a prediction algorithm [17,18].”

14) Figure 1 & 2: there is a typo (“dispnea”)

Authors’ response:

This has been corrected.

6. PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files.

If you choose “no”, your identity will remain anonymous but your review may still be made public.

Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy.

Reviewer #1: No

Reviewer #2: No

Attachment

Submitted filename: Response to Reviewers final.docx

Decision Letter 1

Paola Faverio

2 Sep 2020

Machine learning in predicting respiratory failure in patients with COVID-19 pneumonia - challenges, strengths, and opportunities in a global health emergency

PONE-D-20-16454R1

Dear Dr. Guaraldi,

We’re pleased to inform you that your manuscript has been judged scientifically suitable for publication and will be formally accepted for publication once it meets all outstanding technical requirements.

Within one week, you’ll receive an e-mail detailing the required amendments. When these have been addressed, you’ll receive a formal acceptance letter and your manuscript will be scheduled for publication.

An invoice for payment will follow shortly after the formal acceptance. To ensure an efficient process, please log into Editorial Manager at http://www.editorialmanager.com/pone/, click the 'Update My Information' link at the top of the page, and double check that your user information is up-to-date. If you have any billing related questions, please contact our Author Billing department directly at authorbilling@plos.org.

If your institution or institutions have a press office, please notify them about your upcoming paper to help maximize its impact. If they’ll be preparing press materials, please inform our press team as soon as possible -- no later than 48 hours after receiving the formal acceptance. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact onepress@plos.org.

Kind regards,

Paola Faverio

Academic Editor

PLOS ONE

Additional Editor Comments:

Thank you for addressing all the issues highlighted in the revision process.

Reviewers' comments:

Reviewer's Responses to Questions

Comments to the Author

1. If the authors have adequately addressed your comments raised in a previous round of review and you feel that this manuscript is now acceptable for publication, you may indicate that here to bypass the “Comments to the Author” section, enter your conflict of interest statement in the “Confidential to Editor” section, and submit your "Accept" recommendation.

Reviewer #1: All comments have been addressed

Reviewer #2: All comments have been addressed

**********

2. Is the manuscript technically sound, and do the data support the conclusions?

The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented.

Reviewer #1: Yes

Reviewer #2: Yes

**********

3. Has the statistical analysis been performed appropriately and rigorously?

Reviewer #1: Yes

Reviewer #2: Yes

**********

4. Have the authors made all data underlying the findings in their manuscript fully available?

The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified.

Reviewer #1: Yes

Reviewer #2: Yes

**********

5. Is the manuscript presented in an intelligible fashion and written in standard English?

PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here.

Reviewer #1: Yes

Reviewer #2: Yes

**********

6. Review Comments to the Author

Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters)

Reviewer #1: The authors have responded to all my comments, and the current version of the article in my opinion can be published

Reviewer #2: I would like to thank the Authors for the changes provided. In my opinion, the revised manuscript draft has been improved and is better balanced. I don't have any further comment.

**********

7. PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files.

If you choose “no”, your identity will remain anonymous but your review may still be made public.

Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy.

Reviewer #1: Yes: Catia Cilloniz

Reviewer #2: No

Acceptance letter

Paola Faverio

4 Nov 2020

PONE-D-20-16454R1

Machine learning in predicting respiratory failure in patients with COVID-19 pneumonia - challenges, strengths, and opportunities in a global health emergency

Dear Dr. Guaraldi:

I'm pleased to inform you that your manuscript has been deemed suitable for publication in PLOS ONE. Congratulations! Your manuscript is now with our production department.

If your institution or institutions have a press office, please let them know about your upcoming paper now to help maximize its impact. If they'll be preparing press materials, please inform our press team within the next 48 hours. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information please contact onepress@plos.org.

If we can help with anything else, please email us at plosone@plos.org.

Thank you for submitting your work to PLOS ONE and supporting open access.

Kind regards,

PLOS ONE Editorial Office Staff

on behalf of

Dr. Paola Faverio

Academic Editor

PLOS ONE

Associated Data

    This section collects any data citations, data availability statements, or supplementary materials included in this article.

    Supplementary Materials

    S1 Fig

    (PDF)

    S1 File

    (DOCX)

    Attachment

    Submitted filename: Response to Reviewers final.docx

    Data Availability Statement

    All relevant data are within the manuscript and its Supporting Information files.


    Articles from PLoS ONE are provided here courtesy of PLOS

    RESOURCES