Predicting treatment outcomes following an exacerbation of airways disease

Andreas Halner; Sally Beer; Richard Pullinger; Mona Bafadhel; Richard E K Russell

doi:10.1371/journal.pone.0254425

. 2021 Aug 20;16(8):e0254425. doi: 10.1371/journal.pone.0254425

Predicting treatment outcomes following an exacerbation of airways disease

Andreas Halner ¹, Sally Beer ², Richard Pullinger ², Mona Bafadhel ¹, Richard E K Russell ^1,^3,^*

Editor: Eugene Demidenko⁴

PMCID: PMC8378754 PMID: 34415919

Abstract

Background

COPD and asthma exacerbations result in many emergency department admissions. Not all treatments are successful, often leading to hospital readmissions.

Aims

We sought to develop predictive models for exacerbation treatment outcome in a cohort of exacerbating asthma and COPD patients presenting to the emergency department.

Methods

Treatment failure was defined as the need for additional systemic corticosteroids (SCS) and/or antibiotics, hospital readmissison or death within 30 days of initial emergency department visit. We performed univariate analysis comparing characteristics of patients either given or not given SCS at exacerbation and of patients who succeeded versus failed treatment. Patient demographics, medications and exacerbation symptoms, physiology and biology were available. We developed multivariate random forest models to identify predictors of SCS prescription and for predicting treatment failure.

Results

Data were available for 81 patients, 43 (53%) of whom failed treatment. 64 (79%) of patients were given SCS. A random forest model using presence of wheeze at exacerbation and blood eosinophil percentage predicted SCS prescription with area under receiver operating characteristic curve (AUC) 0.69. An 11 variable random forest model (which included medication, previous exacerbations, symptoms and quality of life scores) could predict treatment failure with AUC 0.81. A random forest model using just the two best predictors of treatment failure, namely, visual analogue scale for breathlessness and sputum purulence, predicted treatment failure with AUC 0.68.

Conclusion

Prediction of exacerbation treatment outcome can be achieved via supervised machine learning combining different predictors at exacerbation. Validation of our predictive models in separate, larger patient cohorts is required.

Introduction

Exacerbations of COPD and asthma are disruptive to patients’ lives, are associated with increased risk of future exacerbations, increased mortality risk and are a large economic burden; furthermore, each event may require hospital admission [1–3]. Not all patients improve following standard initial treatment (as per NICE guidelines with SCS and/or antibiotics [2, 4]) and thus may require additional treatment with systemic corticosteroids (SCS) and/or antibiotics or even hospital readmission [1, 2]. Although part of guidance, studies demonstrate inconsistent benefits for SCS and/or antibiotics for treating exacerbations of COPD and asthma in which treatment is received in hospital [5–8], with increasing concern that SCS causes harm [9]. Single centre [10] and multi-centre studies [11] have examined the use of the peripheral blood eosinophil at the time of an exacerbation to direct SCS use in patients with COPD exacerbation, but little is known about why some patients admitted to hospital succeed whereas others fail treatment. There is thus an urgent need to develop predictive tools to identify which patients at exacerbation will respond successfully to treatment and which patients will not. Such a tool could help identify patients at high risk of treatment failure who require closer monitoring following their presentation to the emergency department (ED), as well as developing biomarker-driven treatment algorithms to optimise patient response to treatment. Since ED presentations are not treated by pulmonologists, a tool to help guide treatment of response in a personalised medicine way is needed for specialists and non-specialists.

In this study, we investigate factors, in patients that attend ED with an exacerbation of asthma or COPD, which may be associated with treatment failure and physician treatment prescription decision. We then use a data-driven approach to develop a multivariate supervised learning algorithm to predict treatment failure of asthma and COPD patients admitted with an exacerbation, as well as to predict the physician decision to prescribe SCS at exacerbation.

Methods

Participants

Participants were recruited if they had a patient-reported primary or secondary care diagnosis of asthma or COPD and presented to the ED of the John Radcliffe Hospital, a large teaching hospital in Oxfordshire, with an exacerbation of COPD or asthma. Exacerbations of COPD and asthma were defined and treated as per GOLD [12] and BTS respectively [13]. Participants with a current history of active pulmonary tuberculosis or current primary malignancy or with any other clinically relevant lung disease judged to be the primary diagnosis were excluded. Any alternative causes for non-exacerbation related increase in symptoms were also excluded, including but not limited to pneumonia, pulmonary embolism, pneumothorax or primary ischemic event. All participants provided informed written consent and the study was approved by the North West London Research Ethics Committee (REC: 15/LO/2119).

Study design

Observational and routinely collected clinical data were prospectively collected for a study duration of 12 months. Study measurements were made on the day of exacerbation termed D0 and on the 1^st, 5^th day (or day of discharge) and 30^th day after exacerbation termed D1, D5 and D30.

Measurements

Data collection included demographics and medication history as well as medical history data such as past asthma or COPD diagnosis. Participant symptoms, health status and quality of life were assessed using participant reported outcome measures including Visual Analogue Scale (VAS) [14], Medical Research Council dyspnoea scale [15], Hospital Anxiety and Depression scale [16] and EuroQol 5D [17]. Physiological measurements taken were the resting oxygen saturations, heart rate and respiratory rate. Venous blood samples were collected and analysed to characterise participants’ inflammatory phenotype; these measurements included peripheral blood eosinophils, C-reactive protein (CRP), biochemistry and glucose.

Statistical analyses

Only participants with complete D30 follow-up data were considered for statistical analysis performed using the programming language ‘R’ [18]. Numerical data were first assessed for normality using the shapiro-wilk test. Normally-distributed data are shown as mean (range) whereas non-normally distributed data are shown as median (interquartile range (IQR)). Treatment failure was defined as a new hospital readmission and/or new and additional treatment (antibiotics and/or SCS) or death between the day of discharge and D30. For numerical variables, the t-test or wilcox test were used for comparing two groups depending on whether the data were normally or non-normally distributed respectively. Data transformations to non-normally distributed variables was not performed. When comparing more than two groups, either the one-way anova test (and Tukey post hoc test) or the kruskal-wallis test (and post-hoc Kruskal Dunn test) was used, depending on whether the normality assumption was met. The chi-squared test was used for binary or categorical data. The relationship between exacerbation VAS symptoms and peripheral blood biomarkers was calculated using Spearman’s rho (r_s). Exploratory univariate analysis of variables at the time of exacerbation was performed to identify differences in characteristics of participants based on exacerbation treatment and identification of variables at D0 variables which differ between treatment failure versus treatment success. Multivariate random forest models were drawn up to predict treatment failure (defined as above) and SCS clinical prescription (see S1 Appendix for more details).

The random forests were then trained on subsets of the variables as part of a feature elimination procedure with leave-one-out cross validation which allows validation of the model performance (see S1 Appendix for full statistical methods) [19, 20].

Results

Out of 104 participants who entered the study, there were 83 with D30 follow-up data. Complete data were available for analysis in 81 participants. A diagnosis of asthma and COPD was found in 59 (73%) and 22 (27%) participants respectively. Treatment allocation showed that at the time of an exacerbation 31 participants (38%) received both SCS and antibiotics; 33 participants (41%) received SCS alone; 6 participants (7%) received antibiotics alone; and 11 participants (14%) received neither. The demographics and at exacerbation characteristics of the asthma and COPD participants are shown in Table 1. The participants with COPD were older (median (IQR) age 67 (60–72) years compared to participants with asthma 41 (28–55) years, p<0.01) and at exacerbation had a lower % oxygen saturation (median oxygen saturation in COPD 93% versus median in asthma 95%, p<0.01). At exacerbation, there was no difference in VAS symptoms of cough, breathlessness, sputum production or purulence between asthma or COPD at the time of exacerbation nor peripheral blood counts. VAS symptoms and peripheral blood inflammatory mediators at exacerbation did not demonstrate a correlation.

Table 1. Characteristics of the asthma and COPD participants at exacerbation.

Characteristic	Primary Respiratory Diagnosis		P-value
Characteristic	Asthma (n = 59)	COPD (n = 22)	P-value
Male, n (%)	20 (34)	14 (64)	0.02
Current smokers, n (%)	15 (25)	6 (27)	<0.001
Ex-smokers, n (%)	18 (31)	16 (73)
Never smoker, n (%)	26 (44)	0 (0)
Number taking ICS, n (%)	33 (56)	7 (32)	0.05
Increased wheeze at exacerbation, n (%)	51 (86)	19 (86)	0.99
Age (years)	41 (28–55)	67 (60–72)	<0.001
Pack year history	0.3 (0.0–12.4)	33.9 (16.0–52.5)	<0.001
Number of hospital admissions in previous 12 months	0 (0–1)	0 (0–1)	0.68
Oxygen saturation (%)	95 (94–97)	93 (91–95)	<0.001
VAS cough (mm)	55 (26–72)	42 (25–53)	0.22
VAS breathlessness (mm)	70 (29–84)	76 (44–97)	0.20
VAS sputum production (mm)	19 (2–48)	21 (1–41)	0.96
VAS sputum purulence (mm)	5 (0–48)	23 (0–53)	0.29
Leucocytes, (x10⁹cells/L)	10.8 (8.9–12.2)	10.7 (9.1–15.5)	0.25
Neutrophils (x10⁹cells/L)	7.9 (5.4–9.5)	7.9 (5.9–12.2)	0.19
Eosinophils (x10⁹cells/L)	0.2 (0.0–0.4)	0.1 (0.0–0.3)	0.65
CRP (mg/L)	7.2 (1.7–23.6)	17.3 (1.8–58.1)	0.14

Open in a new tab

Definition of abbreviations: ICS = inhaled corticosteroids; CRP = C-reactive protein.

Measures of central tendency and spread: all variables are presented as median (IQR) or as number (%) of instances. Comparisons between asthma and COPD are made using chi-squared or wilcox test as appropriate. P-values are reported to 2dp unless the p-value is less than 0.001, in which case it is reported as <0.001.

Characteristics according to D0 treatment

Levels of symptoms and peripheral blood inflammatory mediators compared between participants prescribed both SCS and antibiotic, SCS alone, antibiotics alone, or neither treatment for the exacerbation at D0 are summarised in Table 2. Participants who received both SCS and antibiotics, in addition to those who received no treatment had a higher peripheral blood neutrophil count compared to participants who received SCS alone (post-hoc Kruskal Dunn test p<0.01 and p = 0.01 respectively). Eosinophilic inflammation was highest in those prescribed SCS alone as indicated by a higher peripheral blood eosinophil count and percentage (post-hoc Kruskal Dunn test p < 0.01). Participants who received antibiotics had higher D0 VAS sputum symptoms (see Table 2).

Table 2. Comparison of characteristics of participants given SCS only versus participants given antibiotics only versus participants given both SCS and antibiotics versus participants given neither SCS nor antibiotics.

Characteristic	D0 Antibiotics and SCS Treatment				P-value
	Both	SCS Only	Antibiotics Only	Neither
	(n = 31)	(n = 33)	(n = 6)	(n = 11)
Male n (%)	11 (35)	15 (45)	3 (50)	5 (45)	0.82
Current smokers, n (%)	8 (26)	8 (24)	1 (17)	4 (36)	0.87
Ex-smokers, n (%)	18 (58)	12 (36)	2 (33)	2 (18)	0.29
Never smoker, n (%)	5 (16)	13 (39)	3 (50)	5 (45)	0.24
Asthma, n (%)	20 (65)	26 (79)	5 (83)	8 (73)	0.91
COPD, n (%)	11 (35)	7 (21)	1 (17)	3 (27)	0.69
Number taking ICS, n (%)	13 (42)	18 (55)	4 (67)	5 (45)	0.61
Increased wheeze at exacerbation, n (%)	29 (94)	29 (88)	4 (67)	8 (73)	0.16
Age (years)*	55 (39–69)	45 (28–67)	54 (46–76)	49 (34–58)	0.23
Pack year history*	14 (2–25)	3 (0–18)	0 (0–2)	1 (0–22)	0.22
Number of hospital admissions in previous 12 months*	0 (0–2)	0 (0–1)	0 (0–1)	0 (0–5)	0.90
Oxygen saturation (%)˜	94 (90–98)	95 (86–100)	96 (91–100)	96 (93–100)	0.36
VAS cough (mm)*	55 (30–72)	36 (16–70)	62 (50–78)	30 (53–67)	0.29
VAS breathlessness (mm)*	76 (50–87)	65 (26–84)	70 (61–89)	70 (0–88)	0.78
VAS sputum production (mm)*	28 (8–53)	13 (0–25)	31 (24–75)	2 (0–23)	0.01
VAS sputum purulence (mm)*	32 (5–61)	2 (0–21)	17 (0–45)	1 (0–20)	0.04
Leucocytes (x10⁹cells/L)*	11.0 (9.2–14.1)	9.7 (8.1–12.0)	11.9 (10.6–14.0)	11.9 (10.7–13.9)	0.16
Neutrophils (x10⁹cells/L)*	8.9 (6.0–11.0)	6.0 (4.5–8.2)	8.1 (6.6–9.0)	8.1 (7.7–11.1)	0.03
Neutrophil percentage*	78.1 (68.6–87.8)	65.4 (55.4–75.1)	73.6 (62.6–77.7)	75.0 (66.5–83.6)	0.01
Eosinophils (x10⁹cells/L)*	0.1 (0.0–0.3)	0.3 (0.2–0.5)	0.0 (0.0–0.1)	0.1 (0.1–0.1)	0.01
Eosinophil percentage*	1.2 (0.2–3.1)	3.0 (1.6–5.5)	0.4 (0.1–1.1)	0.9 (0.4–1.1)	0.00
CRP (mg/L)*	19.3 (4.6–42.8)	5.4 (1.0–16.6)	13.4 (1.7–19.8)	3.4 (1.3–17.9)	0.06

Open in a new tab

Definition of abbreviations: ICS = inhaled corticosteroids; CRP = C-reactive protein.

Measures of central tendency and spread: variables marked with the * symbol are presented as median (IQR); variables marked with the ˜ symbol are presented as mean (range). Comparisons between treatment groups are made using chi-squared, kruskal-wallis or one-way anova test as appropriate. P-values are reported to 2dp unless the p-value is less than 0.001, in which case it is reported as <0.001.

Predicting physician SCS prescription decision

SCS were prescribed in 64 (79%) participants. Table 3 summarises characteristics of participants given SCS or not (for which p < 0.1). The remaining characteristics are summarised in S1 Table in S1 Appendix. Participants given SCS had a higher peripheral blood eosinophil percentage compared to participants not given SCS (p < 0.01). Indicators of respiratory disease severity (blood pCO2 and oxygen saturation) were worse in participants given SCS and participants were more likely to report wheeze. Random forest models were developed to determine which combination of biological and symptom variables are the best predictors of the physician decision to prescribe SCS (see S2 Table in S1 Appendix). The variables from the random forest multivariate model showed that the presence of increased wheeze at exacerbation together with blood eosinophil percentage were the best. A random forest model using just these two variables achieved an area under receiver operating characteristic curve (AUC) of 0.69 for predicting SCS prescription (see S3 Table in S1 Appendix for AUCs of the different models).

Table 3. Comparison of D0 exacerbation characteristics of patients given SCS versus patients not given SCS for which p < 0.1 (see S1 Appendix for all other characteristics evaluated).

Characteristic	Exacerbation
	n = 81 Events
	SCS	No SCS	P-value
	n = 64	n = 17	P-value
LAMA, n (%)	13 (20)	10 (59)	0.00
PPI, n (%)	4 (6)	5 (29)	0.01
Eosinophils %*	1.9 (0.4–4.8)	0.8 (0.2–1.1)	0.01
Increased wheeze at exacerbation, n (%)	58 (91)	12 (71)	0.03
Blood pCO2 (kPa)*	5.7 (5.1–6.1)	4.8 (4.5–5.7)	0.05
Total number of ITU admissions in previous 12 months*	0 (0–0)	0 (0–1)	0.05
SCS in week prior to exacerbation admission, n (%)	33 (52)	13 (76)	0.07
Leucocytes (x10⁹cells/L)*	10.2 (8.8–12.8)	11.9 (10.6–14.0)	0.08
Oxygen saturation (%)˜	95 (86–100)	96 (91–100)	0.09

Open in a new tab

Definition of abbreviations: LAMA = Long-acting muscarinic antagonist; PPI = proton pump inhibitor; ITU = intensive therapy unit.

Measures of central tendency and spread: variables marked with the * symbol are presented as median (IQR); variables marked with the ˜ symbol are presented as mean (range). Comparisons between SCS and no SCS group are made using chi-squared, wilcox test or t-test as appropriate. P-values are reported to 2dp unless the p-value is less than 0.001, in which case it is reported as <0.001.

Predicting a treatment failure

In total, there were 43 (53%) treatment failures. Participants who failed treatment were likely to have had a history of exacerbations, a higher symptom burden of dyspnoea, sputum production and purulence at exacerbation and were less likely to be taking inhaled COPD treatment. Participants prescribed SCS were also more likely to have a treatment failure. Table 4 summarises the characteristics of participants who did and did not have a treatment failure for which p < 0.1. The full comparison of characteristics of these participants is summarised in S4 Table in S1 Appendix. Table 5 shows the random forest variable importance scores for the multivariate model containing all variables which passed the univariate p<0.1 filter. The best model used 11 variables and achieved an AUC of 0.81 (see S5 Table in S1 Appendix for AUCs of the different random forest models using different combinations and total numbers of variables) for predicting treatment failure. VAS breathlessness and VAS sputum purulence at exacerbation contribute most to the predictive performance. A random forest model using just these two variables predicted treatment failure with AUC 0.68.

Table 4. Comparison of D0 exacerbation characteristics of patients who failed treatment with those who succeeded treatment which p < 0.1 (see S1 Appendix for all other characteristics evaluated).

Characteristic	Exacerbation
	n = 81 Events
	Treatment Failure, n = 43	Treatment Success, n = 38	P-value
VAS sputum production (mm)*	28 (10–56)	8 (0–25)	0.00
VAS dyspnoea (mm)*	78 (60–89)	55 (4–78)	0.00
Number of exacerbations associated with increased sputum production, n (%)	28 (65)	13 (34)	0.01
Number of unscheduled primary care and emergency department visits in previous 12 months*	2 (1–5)	1 (0–3)	0.01
Number taking ICS, n (%)	15 (35)	25 (66)	0.01
VAS sputum purulence (mm)*	40 (0–68)	3 (0–15)	0.01
Number of exacerbations associated with increased sputum purulence, n (%)	22 (51)	10 (26)	0.02
SCS at exacerbation, n (%)	38 (88)	26 (68)	0.03
Oxygen saturation, n (%)˜	94 (86–100)	96 (91–100)	0.03
SABA, n (%)	32 (74)	35 (92)	0.04
LAMA, n (%)	8 (19)	15 (39)	0.04
PPI, n (%)	2 (5)	7 (18)	0.05
VAS cough (mm)˜	54 (0–99)	43 (0–100)	0.08

Open in a new tab

Definition of abbreviations: ICS = inhaled corticosteroids; SABA = short-acting beta agonist; LAMA = Long-acting muscarinic antagonist; PPI = proton pump inhibitor.

Measures of central tendency and spread: variables marked with the * symbol are presented as median (IQR); variables marked with the ˜ symbol are presented as mean (range). Comparisons between treatment failure and treatment success are made using chi-squared, wilcox test or t-test as appropriate. P-values are reported to 2dp unless the p-value is less than 0.001, in which case it is reported as <0.001.

Table 5. Random forest importance scores of variables in multivariate treatment failure prediction models for all participants.

Name of Successive Variable	Importance Score of Successive Variable
VAS breathlessness	11.2
VAS sputum purulence	8.3
LAMA	6.5
Number of unscheduled primary care and emergency department visits in previous 12 months	5.6
ICS use	4.9
VAS sputum production	4.2
PPI use	4.2
EuroQol mobility	4.1
Oxygen saturation	3.8
SABA use	3.3
SCS at exacerbation	2.7
Exacerbation associated with increased sputum purulence	1.1
Exacerbation associated with increased sputum production	1.0
EuroQol self care	0.8
Gender	0.6
EuroQol usual activity	-0.5
VAS cough	-1.0

Open in a new tab

Definition of abbreviations: LAMA = Long-acting muscarinic antagonist; ICS = inhaled corticosteroids; PPI = proton pump inhibitor; SABA = short-acting beta agonist.

Models shown range from those using just a single variable up to those with the full subset which passed the univariate analysis filter of p < 0.1. Variable importance scores were calculated as outlined in S1 Appendix Statistical Methods. To illustrate the interpretation of the variable importance scores, consider the example of VAS breathlessness. The VAS breathlessness random forest importance score of 11.2 indicates that classification accuracy would drop by 11.2% if VAS breathlessness is omitted from the classification model.

Discussion

In this study, we have described the characteristics of patients with asthma and COPD who present to the ED with an exacerbation. We have shown that treatment failure with SCS and/or antibiotic therapy occurs in approximately 50% and that symptoms of breathlessness and sputum purulence appear be good predictors of a treatment failure following an exacerbation of airways disease. In addition to this, the clinician decision to prescribe SCS is related to several factors. Increased wheeze at exacerbation and peripheral blood eosinophil percentage together predict SCS prescription at the onset of an exacerbation of asthma and/or COPD, although it is possible that increased wheeze at exacerbation is a confounder for the relationship between peripheral blood eosinophil percentage and SCS prescription.

In this study we have used random forest analyses to identify factors associated with exacerbation treatment failure. We showed that patients who went on to fail treatment had more pronounced breathlessness, sputum production and sputum purulence at exacerbation, indicative of symptom burden. Exacerbations are symptom-defined events [1] and our findings illustrate this. Additionally our results potentially allow for quantitative qualification of the symptoms measured at the onset of the event that are associated with exacerbations that fail to respond to management at the onset. Furthermore, patients who were on inhaled therapy (corticosteroids, short-acting beta agonists and/or long-acting muscarinic antagonists) were less likely to have a treatment failure following an exacerbation, reiterating that undertreated COPD is associated with a higher chance of a patient requiring further healthcare utilisation [21]. We also showed that the use of PPI appeared to be protective of a treatment failure. This was interesting as gastro-oesphageal reflux disease has previously been shown to be an independent predictor of exacerbations [22, 23].

Our random forest analysis to predict treatment failure showed that a random forest model using a combination of breathlessness and sputum purulence exacerbation symptom VAS scores could effectively predict treatment failure for patients irrespective of the type of exacerbation treatment patients were given. Our finding that breathlessness at exacerbation is a useful predictor of treatment failure compliments findings of other studies. Breathlessness is already known to be a predictor of 5 year mortality for COPD patients [24] whilst the Medical Research Council Dyspnoea Scale is a predictor of both hospital mortality and 28-day exacerbation readmission [25]. Similarly, dyspnoea has been shown to be associated with exacerbation relapse risk in patients with acute exacerbations of COPD [26]. Our finding that sputum purulence is a predictor of exacerbation treatment failure is interesting in the context of GOLD and NICE guidance that antibiotics should be given at exacerbation in patients with increased sputum purulence since this may reduce exacerbation relapse and treatment failure [1, 27]. The relatively high prescription of antibiotics for asthma exacerbations in our study is not an uncommon occurrence, as has previously been shown [28]. This is likely due to patients not being seen by specialists in ED and reflects clinical practice, further emphasising the importance of developing tools to assist non-specialists in ED. In our study it was observed that patients receiving antibiotics at exacerbation did indeed have significantly greater sputum purulence compared to patients not receiving antibiotics. Therefore, it is unlikely that treatment failures in our study resulted from inappropriate or non- prescription of antibiotics at exacerbation. Rather, our finding that higher dyspnoea and sputum purulence are predictive of treatment failure may simply reflect that patients with greater disease burden and more severe exacerbations in terms of symptomatic presentation are more likely to fail treatment. It may also be the case that the patients with the highest symptom load always have a high symptom load, so the threshold to declare this is reached more frequently and treatment may not lower their symptoms enough for them to drop below their threshold. Our conclusion that treatment failure can be predicted based on exacerbation symptomatological presentation reaffirms the need for clinicians to pay close attention to patient symptom presentation at exacerbation. Furthermore, the ease of measuring patient symptoms through the VAS score would make treatment failure prediction models like those in our study easy to implement in clinical practice.

It is noteworthy that a high symptom burden for VAS sputum production and sputum purulence in addition to CRP were characteristic of patients given antibiotics, although CRP did not reach statistical significance. It has been shown that using CRP as a biomarker to direct antibiotic treatment of severe COPD exacerbations is effective in achieving good clinical treatment outcomes as well as reducing antibiotic prescription [29]. The latter is especially important given the concerns associated with excessive antibiotic use for treating exacerbations promoting antibiotic resistance [30]. The only biological exacerbation characteristic which differed between patients given SCS and patients not given SCS was the degree of eosinophilic inflammation. Blood results would not have been available for clinicians prior to treatment initiation. This is an interesting finding, as it suggests that there is something about eosinophilic inflammation that drives a clinician to prescribe SCS. In investigating this further, our univariate analysis, showed that prescription of a LAMA and PPI, was associated with reduced SCS prescription, whilst an elevated pCO₂, and the presence of wheeze was associated with SCS prescription. Random forest models revealed that blood eosinophil percentage and increased wheeze at exacerbation were predictors of SCS prescription. Eosinophilic inflammation is related to increased luminal airway oedema [31] and in asthma post-mortem narrowed airways [32], which could suggest that eosinophilic inflammation is related to wheeze and a greater degree of clinical severity of the presentation. In the context of our study, this may explain the greater predominance of wheeze at exacerbation as well as eosinophilic inflammation for those patients given SCS. Studies have shown that patients with a high peripheral blood eosinophil count respond better to SCS than patients with a low blood eosinophil count [10] and our finding suggests the importance of measuring eosinophils in all exacerbations of airways disease.

Our study has some limitations. Firstly, combination of asthma and COPD may make interpretation difficult. However, as is standard practice, often there is no access to lung function in ED from primary care. In addition, where monitoring is now impacted by the absence of spirometry in the community setting due to COVID, it is prudent to make the assumption that a clear-cut diagnosis of asthma or COPD is not always available. The combination of asthma and COPD can thus be seen as an advantage rather than a limitation. A second potential limitation is that we did not perform SCS prescription and treatment failure analyses for asthma compared to COPD patients. This is because out of the 81 patients, only 22 had a primary diagnosis of COPD, which would be too small a sample size for separate analysis. However, whether the primary diagnosis was asthma or COPD is a variable considered in our SCS prescription and treatment failure prediction analysis and this variable is not selected in the best random forest models. Hence, we conclude that whether the diagnosis is asthma versus COPD is not related to SCS prescription and treatment failure in our study cohort. We note that a third limitation of our study is the lack of external validation. However, our leave-one-out cross-validation strategy already enabled validation of different models to be performed in test subjects. In the context of our small sample size, this is superior compared to splitting the dataset into one training and one testing portion which would waste data otherwise available for training [20]. Furthermore, our choice of random forest as the supervised learning approach was strategic to reduce the potential of over-fitting [19]. We envisage that in additional future clinical validation studies it will be more reliable to use the variables in their original form rather than in a transformed form for input into the multivariate classifier being assessed in a validation study. Therefore, in our univariate analysis we compared different types of p-values including chi-squared, t-test and Wilcox test. The mixture of continuous, discrete, binary and categorical variables in the study made comparing at least two different types of p-values unavoidable. For our univariate analysis we did not correct for multiple comparisons since the univariate analysis was purely exploratory in nature. Nevertheless, we note that the data-driven, assumption-free machine learning approach in our study is a key strength. It should be noted that when narrowing down the size of variable subsets in final multivariate models for treatment failure prediction, we deliberately used the random forest variable importance score and feature elimination procedure rather than relying on potentially more easily interpretable univariate p-values or AUCs of single variable models to indicate variable importance. Ordering variables according to the corresponding p values from the initial univariate analysis would be mathematically inaccurate since: 1) the p-value would only indicate univariate importance of a variable which may be very different from the importance of the variable in a multivariate context [20]; 2) the p-values relate to the hypothesis test for a difference between the treatment failure and treatment success groups rather than being a direct measure of the ability of the variable to classify [20] a patient as being in the treatment failure versus treatment success group. In addition, it is mathematically best to use a variable ranking system directly related to the particular classifier model being used since this provides more information regarding the generalisability of the variable’s importance for classification using the relevant model [20].

To the best of our knowledge, our study is the first to assess a combination of VAS-based symptoms, biomarkers, clinical characteristics and demographics as predictors of treatment failure in a mixed cohort of hospitalised asthma and COPD patients.

Conclusion

We have shown that over half of all asthma and COPD patients admitted to ED for their exacerbation require additional major medication and/or readmission to the emergency department within 30 days of their exacerbation. We have also shown that prescription of SCS at exacerbation by physicians may be related to presentation of wheeze and may be driven by eosinophilic inflammation. Furthermore, breathlessness and sputum purulence are useful for predicting whether additional major medication and/or readmission to the emergency department within 30 days of exacerbation will occur.

Supporting information

S1 Appendix. Predicting treatment outcomes following an exacerbation of airways disease appendix.

(DOCX)

Click here for additional data file.^{(65KB, docx)}

S1 Dataset

(CSV)

Click here for additional data file.^{(25.6KB, csv)}

Acknowledgments

The authors thank all the research volunteers who participated in the study.

Data Availability

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

Funding Statement

The research was supported by the National Institute for Health Research (NIHR) Oxford Biomedical Research Centre (BRC). The views expressed are not necessarily those of the NHS, the NIHR or the Department of Health. The NIHR/NHS/Department of Health had no role in study design, data collection and analysis, decision to publish or preparation of the manuscript.

References

1.Vogelmeier CF, Criner GJ, Martinez FJ, et al. Global strategy for the diagnosis, management, and prevention of chronic obstructive lung disease 2017 report. Am J Respir Crit Care Med. 2017; 195: 557–582. doi: 10.1164/rccm.201701-0218PP [DOI] [PubMed] [Google Scholar]
2.Chronic obstructive pulmonary disease in over 16s: diagnosis and management | Guidance | NICE. (2010). [PubMed]
3.Greening NJ, Williams JEA, Hussain SF, et al. An early rehabilitation intervention to enhance recovery during hospital admission for an exacerbation of chronic respiratory disease: Randomised controlled trial. BMJ. 2014; 349(7967): G4315. doi: 10.1136/bmj.g4315 [DOI] [PMC free article] [PubMed] [Google Scholar]
4.Scenario: Acute exacerbation of asthma. 2020. Available from: https://cks.nice.org.uk/topics/asthma/management/acute-exacerbation-of-asthma/.
5.Walters JAE, Tan DJ, White CJ, Gibson PG, Wood-Baker R, Walters EH. Systemic corticosteroids for acute exacerbations of chronic obstructive pulmonary disease. Cochrane Database Syst Rev. 2014; (9):CD001288. doi: 10.1002/14651858.CD001288.pub4 [DOI] [PMC free article] [PubMed] [Google Scholar]
6.Vollenweider DJ, Frei A, Steurer-Stey CA, Garcia-Aymerich J, Puhan MA. Antibiotics for exacerbations of chronic obstructive pulmonary disease. Cochrane Database Syst Rev. 2018; 10(10):CD010257. doi: 10.1002/14651858.CD010257.pub2 [DOI] [PMC free article] [PubMed] [Google Scholar]
7.Phipatanakul W, Mauger DT, Sorkness RL, et al. Effects of Age and Disease Severity on Systemic Corticosteroid Responses in Asthma. Am J Respir Crit Care Med. 2017; 95(11):1439–1448. doi: 10.1164/rccm.201607-1453OC [DOI] [PMC free article] [PubMed] [Google Scholar]
8.Normansell R, Sayer B, Waterson S, Dennett EJ, Del Forno M, Dunleavy A. Antibiotics for exacerbations of asthma (Review). Cochrane Database Syst Rev. 2018; (6): CD002741. doi: 10.1002/14651858.CD002741.pub2 [DOI] [PMC free article] [PubMed] [Google Scholar]
9.Waljee A, Rogers M, Lin P, et al. Short term use of oral corticosteroids and related harms among adults in the United States: population based cohort study. BMJ 2017;357:j1415. doi: 10.1136/bmj.j1415 [DOI] [PMC free article] [PubMed] [Google Scholar]
10.Bafadhel M, McKenna S, Terry S, et al. Blood eosinophils to direct corticosteroid treatment of exacerbations of chronic obstructive pulmonary disease: A randomized placebo-controlled trial. Am. J. Respir. Crit. Care Med. 2012; 186(1): 48–55. doi: 10.1164/rccm.201108-1553OC [DOI] [PMC free article] [PubMed] [Google Scholar]
11.Sivapalan P, Lapperre TS, Janner J, et al. Eosinophil-guided corticosteroid therapy in patients admitted to hospital with COPD exacerbation (CORTICO-COP): a multicentre, randomised, controlled, open-label, non-inferiority trial. Lancet Respir. Med. 2019; 7(8), 699–709. doi: 10.1016/S2213-2600(19)30176-6 [DOI] [PubMed] [Google Scholar]
12.Vestbo J, Hurd SS, Agusti AG, et al. Global strategy for the diagnosis, management, and prevention of chronic obstructive pulmonary disease GOLD executive summary. Am J Respir Crit Care Med. 2013; 187:347–365. doi: 10.1164/rccm.201204-0596PP [DOI] [PubMed] [Google Scholar]
13.British guideline on the management of asthma. (2014). [Google Scholar]
14.Brightling CE, Monterio W, Green RH, et al. Induced sputum and other outcome measures in chronic obstructive pulmonary disease: Safety and repeatability. Respir. Med. 2001; 95(12): 999–1002. doi: 10.1053/rmed.2001.1195 [DOI] [PubMed] [Google Scholar]
15.Fletcher CM, Clifton M, Fairbairn AS, et al. Standardised questionnaire on respiratory symptoms: a statement prepared and approved by the MRC Committee on the Aetiology of Chronic Bronchitis (MRC breathlessness score). BMJ. 1960; 2(5213): 1665. 13688719 [Google Scholar]
16.Zigmond AS, Snaith RP. The Hospital Anxiety and Depression Scale. Acta Psychiatr. Scand. 1983; 67(6): 361–370. doi: 10.1111/j.1600-0447.1983.tb09716.x [DOI] [PubMed] [Google Scholar]
17.Group EuroQol. EuroQol—a new facility for the measurement of health-related quality of life. Health Policy. 1990; 16(3): 199–208. doi: 10.1016/0168-8510(90)90421-9 [DOI] [PubMed] [Google Scholar]
18.R Core Team (2019). R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. URL https://www.R-project.org/. [Google Scholar]
19.Breiman L. Random forests. Mach. Learn. 2001; 45(1): 5–32. [Google Scholar]
20.Kuhn M, Johnson K. Applied Predictive Modeling. New York: Springer; 2013. [Google Scholar]
21.Ding B, Small M, Bergström G, Holmgren U. COPD Symptom Burden: Impact on Health Care Resource Utilization, and Work and Activity Impairment. Int J Chron Obstruct Pulmon Dis. 2017; 12: 677–89. doi: 10.2147/COPD.S123896 [DOI] [PMC free article] [PubMed] [Google Scholar]
22.Hurst JR, Vestbo J, Anzueto A, et al. Susceptibility to Exacerbation in Chronic Obstructive Pulmonary Disease. N Engl J Med. 2010; 363: 1128–138. doi: 10.1056/NEJMoa0909883 [DOI] [PubMed] [Google Scholar]
23.Sakae TM, Pizzichini MMM, Teixeira PJZ, Maurici Da Silva R, Trevisol DJ, Pizzichini E. Exacerbations of COPD and Symptoms of Gastroesophageal Reflux: A Systematic Review and Meta-analysis. J Bras Pneumol. 2013; 39(3): 259–71. doi: 10.1590/S1806-37132013000300002 [DOI] [PMC free article] [PubMed] [Google Scholar]
24.Nishimura K, Izumi T, Tsukino M, Oga T. Dyspnea is a better predictor of 5-year survival than airway obstruction in patients with COPD. Chest. 2002; 121(5): 1434–1440. doi: 10.1378/chest.121.5.1434 [DOI] [PubMed] [Google Scholar]
25.Steer J, Norman EM, Afolabi OA, Gibson GJ, Bourke SC. Dyspnoea severity and pneumonia as predictors of in-hospital mortality and early readmission in acute exacerbations of COPD. Thorax. 2012; 67(2): 117–21. doi: 10.1136/thoraxjnl-2011-200332 [DOI] [PubMed] [Google Scholar]
26.Mantero M, Rogliani P, Di Pasquale M, et al. Acute exacerbations of COPD: Risk factors for failure and relapse. Int J Chron Obstruct Pulmon Dis. 2017; 12: 2687–2693. doi: 10.2147/COPD.S145253 [DOI] [PMC free article] [PubMed] [Google Scholar]
27.Chronic obstructive pulmonary disease (acute exacerbation): antimicrobial prescribing (NG114) (2018).
28.Bafadhel M, Clark T, Reid C, et al. Procalcitonin and C-reactive protein in hospitalized adult patients with community-acquired pneumonia or exacerbation of asthma or COPD. Chest. 2010; 139(6):1410–1418. doi: 10.1378/chest.10-1747 [DOI] [PMC free article] [PubMed] [Google Scholar]
29.Prins HJ, Duijkers R, van der Valk P, et al. CRP-guided Antibiotic Treatment in acute exacerbations of COPD admitted to Hospital. Eur Respir J. 2019; 53(5):180201. [DOI] [PubMed] [Google Scholar]
30.Boersma WG. Antibiotics in acute exacerbations of COPD: The good, the bad and the ugly. Eur Respir J. 2012; 40(1): 1–3. doi: 10.1183/09031936.00211911 [DOI] [PubMed] [Google Scholar]
31.Trivedi SG, Lloyd CM. Eosinophils in the Pathogenesis of Allergic Airways Disease. Cell Mol Life Sci. 2007; 64(10): 1269–289. doi: 10.1007/s00018-007-6527-y [DOI] [PMC free article] [PubMed] [Google Scholar]
32.Houston JC, De Navasquez S, Trounce JR. A Clinical and Pathological Study of Fatal Cases of Status Asthmaticus. Thorax. 1953; 8:(3): 207–13. doi: 10.1136/thx.8.3.207 [DOI] [PMC free article] [PubMed] [Google Scholar]

PLoS One. doi: 10.1371/journal.pone.0254425.r001

Decision Letter 0

Eugene Demidenko

2 Feb 2021

PONE-D-20-36305

Predicting Treatment Outcomes Following an Exacerbation of Airways Disease

PLOS ONE

Dear Dr. Halner,

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.

The goal of the paper is to build a parsimonious statistical predictive model for treatment failure of COPD and asthma exacerbation. The topic is important and in general the approach is acceptable. However, several issues must be fixed before the paper can be sent out for full review.

Some page numbers are shown and some not – it makes reading of the hard copy difficult.
Comparing p-values obtained by different methods is cumbersome. Why authors don’t use the traditional t-test, perhaps after normalizing transformation, such as taking log?
It is unclear what P-value in Table 2 comes from. What is compared? I suggest showing p-values in scientific format, like 1.2x10^-4, instead <0.01. Such p-value may help in comparing the strength of predictors.
The notation β=1.72 on page 11 comes from nowhere. The coefficient?
I do not see the ref for NagelKerke’s R². I think using AUC, or the C-statistic, is a more appropriate for this study as the measure of correct classification.
Why R² is blank for univariate models in Table 4?
The table results, such as presented in Table 2, are difficult to read. I suggest boxplots instead. This comment applies to all other results including the ROC curve.
The authors claim that the data are fully available. However neither Excel file nor URL is provided where the raw data can be downloaded from.
Random forest results on importance of predictors in Table 6 are questionable because of lack of interpretation. How to interpret the scores and what they mean? Instead, the variables may be ordered with respect to the p-value or better off with respect to increasing values of AUC from expanding logistic regression. The AUC has clear interpretation as % failure prediction.
I don’t see a good reason to include variables with p-value p>=0.03 in the final predictive model presented in Table 5. Instead I suggest computing the C-statistic (AUC) for expanding set of variables. Then the contribution of each predictor could be easily assessed.
The choice of the final model is not justified well. I suggest BIC or AIC criterion.

Please submit your revised manuscript by Mar 19 2021 11:59PM. 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,

Eugene Demidenko, Ph.D.

Academic Editor

PLOS ONE

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.We note that the grant information you provided in the ‘Funding Information’ and ‘Financial Disclosure’ sections do not match.

When you resubmit, please ensure that you provide the correct grant numbers for the awards you received for your study in the ‘Funding Information’ section.

3.Thank you for stating the following in the Competing Interests section:

"I have read the journal's policy and the authors of this manuscript have the following competing interests:

Mona Bafadhel reports outside the submitted work research grant reports from AZ; honoraria from AZ, Chiesi, and GlaxoSmithKline; and is on the scientific advisory board for AlbusHealth® and ProAxsis®. Richard Russell has received honoraria from AZ, GSK, Boheringer Ingelheim, Chiesi, Cipla and is on the advisory board for AlbusHealth®, has received research funding from Circassia UK and his work is supported by the Oxford NIHR Biomedical Research Centre.

The remaining authors have declared that no competing interests exist."

Please confirm that this does not alter your adherence to all PLOS ONE policies on sharing data and materials, by including the following statement: "This does not alter our adherence to PLOS ONE policies on sharing data and materials.” (as detailed online in our guide for authors http://journals.plos.org/plosone/s/competing-interests). If there are restrictions on sharing of data and/or materials, please state these. Please note that we cannot proceed with consideration of your article until this information has been declared.

Please include your updated Competing Interests statement in your cover letter; we will change the online submission form on your behalf.

Please know it is PLOS ONE policy for corresponding authors to declare, on behalf of all authors, all potential competing interests for the purposes of transparency. PLOS defines a competing interest as anything that interferes with, or could reasonably be perceived as interfering with, the full and objective presentation, peer review, editorial decision-making, or publication of research or non-research articles submitted to one of the journals. Competing interests can be financial or non-financial, professional, or personal. Competing interests can arise in relationship to an organization or another person. Please follow this link to our website for more details on competing interests: http://journals.plos.org/plosone/s/competing-interests

[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. 2021 Aug 20;16(8):e0254425. doi: 10.1371/journal.pone.0254425.r002

Author response to Decision Letter 0

4 Mar 2021

RE: Predicting Treatment Outcomes Following an Exacerbation of Airways Disease

REF: PONE-D-20-36305

To the Editor, Dr Eugene Demidenko,

Thank you very much for your consideration of our manuscript. We have taken the time to thoroughly address the reviewer comments through making the suggested edits or providing a justification of our approach. As requested, we have uploaded a tracked changes version named ‘Revised Manuscript with Track Changes PONE-D-20-36305’ (as well as ‘Revised Supplement with Track Changes - PONE-D-20_36305’) and an unmarked version named ‘Manuscript - PONE-D-20-36305_Clean’ (as well as ‘Supplement - PONE-D-20_36305_Clean’).

Please see our responses below:

Comment 1: “Some page numbers are shown and some not – it makes reading of the hard copy difficult.”

Response 1: Thank you for pointing this out. This has been now addressed.

Comment 2: “Comparing p-values obtained by different methods is cumbersome. Why authors don’t use the traditional t-test, perhaps after normalizing transformation, such as taking log?”

Response 2: Thank you. We have intentionally not applied data transformations since we envisage that in future clinical validation studies it will be more reliable to use the variables in their original form rather than in a transformed form for input into the multivariate classifier being assessed in a validation study. This is important since for future-obtained data one cannot be sure of the distribution (e.g., whether normal or not) and since there is a mixture of continuous, discrete, binary or categorical variables comparing at least two types of p-values (i.e., chi-squared and t-test or Wilcox test) is unavoidable. We have discussed this now further in the main body of the manuscript (see ‘Revised Manuscript with Track Changes PONE-D-20-36305’ ‘Statistical Analyses’ section on pg.6 lines 11-12 and ‘Discussion’ on pg.19 lines 18-23).

Comment 3: “It is unclear what P-value in Table 2 comes from. What is compared? I suggest showing p-values in scientific format, like 1.2x10-4, instead <0.01. Such p-value may help in comparing the strength of predictors.”

Response 3: Thank you this has now been addressed and we have now replaced < 0.01 with more precise p-values via the use of standard form, as suggested by the reviewers. We have added detail to the title of Table 2 to clarify the nature of the comparisons between treatment groups. For the sake of consistency, we have now also used standard form for p-values in all other tables in the ‘Revised Manuscript with Track Changes PONE-D-20-36305’.

Comment 4: “The notation β=1.72 on page 11 comes from nowhere. The coefficient?”

Response 4: Thank you this has now been addressed. We have clarified this by writing “β coefficient” rather than just “β” (see ‘Revised Manuscript with Track Changes PONE-D-20-36305’ pg.12 line 10).

Comment 5: “I do not see the ref for NagelKerke’s R2. I think using AUC, or the C-statistic, is a more appropriate for this study as the measure of correct classification.”

Response 5: Thank you this has now been addressed. For each model in Table 4, instead of the NagelKerke’s R2 we have now reported the AUC in addition to Akaike Information Criterion (AIC), with one exception for the univariate model which uses the binary variable wheeze. In the latter case, the AIC but not the AUC is reported. We have outlined the reporting of AUC and AIC, as well as the lack of AUC reporting for the univariate binary predictor case (with references), in the ‘Revised Manuscript with Track Changes PONE-D-20-36305’ (pg.6 line 23 and pg.7 lines 1-2; pg.12 line 12) and ‘Revised Supplement with Track Changes - PONE-D-20_36305' document (see pg.2 lines 10-16).

Comment 6: “Why R2 is blank for univariate models in Table 4?”

Response 6: Thank you. As per comment 5, the R2 in Table 4 has been replaced with AIC and AUC.

Comment 7: “The table results, such as presented in Table 2, are difficult to read. I suggest boxplots instead. This comment applies to all other results including the ROC curve.”

Response 7: Thank you. We have clarified table 2 for readability. We have not included box-plots as per reviewers suggestion as this would require 56 individual plots. If the editorial discretion is to replace table 2 with box-plots we can upload this.

Comment 8: “The authors claim that the data are fully available. However, neither Excel file nor URL is provided where the raw data can be downloaded from.”

Response 8: The data is fully available following written request. This is clarified in the manuscript acknowledgements (see ‘Revised Manuscript with Track Changes PONE-D-20-36305’ ‘Acknowledgements’ section on pg.22).

Comment 9: “Random forest results on importance of predictors in Table 6 are questionable because of lack of interpretation. How to interpret the scores and what they mean? Instead, the variables may be ordered with respect to the p-value or better off with respect to increasing values of AUC from expanding logistic regression. The AUC has clear interpretation as % failure prediction.”

Response 9: Thank you. In our manuscript we have ordered the random forest variable importance scores to show the reader the hierarchy of importance of the variables as contributing to the predictive ability of the multivariate random forest model. For example, presenting a hierarchy of the p value would indicate the importance of the variable in univariate analysis only; whilst the AUCs for random forest models using different combinations of variables are provided in the supplement S1 Appendix Table 3. To add clarity, we have added in the footnote of Table 6 what the importance scores translate to. We have used the example that the VAS breathlessness random forest importance score of 11.2 indicates that classification accuracy would drop by 11.2% if this was omitted.

Comment 10: “I don’t see a good reason to include variables with p-value p>=0.03 in the final predictive model presented in Table 5. Instead, I suggest computing the C-statistic (AUC) for expanding set of variables. Then the contribution of each predictor could be easily assessed.”

Response 10: Thank you. As described in the supplement S1 Appendix we have included all variables with univariate p values < 0.1 in the initial multivariate model and the selected variables (17 variables met the p < 0.1 univariate filter in our treatment failure prediction analysis) were ranked according to random forest variable importance scoring. We have clarified this further as per response 9. We deliberately used an initial p < 0.1 univariate filter rather than a more limiting threshold (e.g., p <0.05 or the p < 0.03 suggestion made by the reviewer), because the initial filter step serves to include as many variables as possible for consideration in the multivariate model but to remove variables likely to represent noise rather than being truly informative. We have added further details in the statistical appendix (see ‘Revised Supplement with Track Changes - PONE-D-20_36305' pg.2 lines 28-33 and pg.3 lines 1-2) as to why this method was chosen and discussed further other methods (e.g., the C-Statistic as per

Comment 10 and p-values as per Comment 9) and their limitations in the main manuscript (see ‘Revised Manuscript with Track Changes PONE-D-20-36305’ pg.20 lines 2-15).

Comment 11: “The choice of the final model is not justified well. I suggest BIC or AIC criterion.”

Response 11: Thank you. Although these criterions could be used, in our model, neither BIC nor AIC can be applied in the context of an ensemble decision tree-based classifier such as the random forest. Both the number of parameters estimated by a classifier and the maximum value of the likelihood function for a model must be known in order to compute BIC or AIC. The primary reason why we used the random forest classifier for predicting treatment failure in our study rather than using a linear parametric model such as the logistic regression is that random forests can be effective in preventing overfitting, even in a low sample size high dimensional context. We have further discussed in the statistical appendix why we selected this model and why AIC/BIC criterion cannot be applied (see ‘Revised Supplement with Track Changes - PONE-D-20_36305' pg.3 lines 11-31 and pg.11 references 1 and 5).

We look forward to hearing from you.

Kind regards,

Mr Andreas Halner & Professor Mona Bafadhel

on behalf of the co-authors

PLoS One. doi: 10.1371/journal.pone.0254425.r003

Decision Letter 1

Eugene Demidenko

11 Mar 2021

PONE-D-20-36305R1

Predicting Treatment Outcomes Following an Exacerbation of Airways Disease

PLOS ONE

Dear Dr. Halner,

The reviews of your manuscript are on the negative side: the first reviewer suggested revision and the second reviewer suggested rejection. Nevertheless, I give you a chance to resubmit. Addressing their concerns in a point-by-point fashion is imperative. Lack of convincing response will result in the subsequent rejection. Especially important is the critique of the second reviewer who made very important objections from the medical perspective. Feel free to withdraw the paper if you think that such response would be difficult provide.

Please submit your revised manuscript by Apr 25 2021 11:59PM. 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'.

We look forward to receiving your revised manuscript.

Kind regards,

Eugene Demidenko, Ph.D.

Academic Editor

PLOS ONE

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

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: (No Response)

Reviewer #2: (No Response)

**********

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: Partly

Reviewer #2: No

**********

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

Reviewer #1: No

Reviewer #2: I Don't Know

**********

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: No

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: Overall, I don't think the methodology has been sufficiently thought through. To give you the benefit of the doubt, the wide mix of analyses and methods mean that you have not been able to thoroughly describe them. I think your aims are a little unclear - are you just looking for factors associated with SCS prescription and treatment failure, or are you trying to predict them? If the latter, the methodology is inappropriate because you are not reporting or testing the performance in a valid way, and the models are not being developed well. If the former, reporting the AIC and AUC seems unnecessary, as does comparing the results across different numbers of variables included.

I think you need to refine your analysis question, and remove some of the analyses.

• Abstract: SCS acronym used in abstract without introduction

• Introduction: please define the ‘standard initial treatment’

• Methods:

o Were there any participants with asthma-COPD overlap syndrome? This is not mentioned at all.

o It is unclear to me why the methods were not consistent between the two main analyses: SCS prescription and treatment failure. You justify the use of the random forest for the second analysis, but it seems like it would have been appropriate for the first analysis for the same reason.

o Insufficient detail is provided on the construction of the multivariate logistic regression models. Did you try multiple combinations of between two and four variables and then select the best performing? Why was a stepwise approach not used?

o By only allowing the random forest to choose from variable significant in the univariate logistic regression, potential interactions between variables are very limited. The ability to calculate such interactions is a key strength of tree-based algorithms.

o Which implementation of the random forest algorithm was used? What were the hyperparameters?

o Did you look at the analyses stratified by diagnosis? It’s reasonable to assume it might differ between asthma and COPD.

• Results:

o Please change ‘asthmatics’ to ‘people with asthma’

o No p-values appeared to be in bold in any of the tables.

o I disagree with the previous reviewer that presenting the p-values in standard form is better, I think your previous approach was better.

o Table 2 just says LAMA, but table 6 says lama use. Be consistent.

o Table 2 – can you group the characteristics by type (continuous, binary etc) for ease of reading?

o Please define acronyms used in tables in the notes underneath them

• Discussion:

o You should not be defining acronym for the first time in the discussions when it has been used throughout, such as PPI.

o Grammatical error: “so the threshold to declare this, is reached more frequently”

o Overly strong assumption of causal relationship between eosinophilic inflammation and SCS prescribing. It seems very possible that this is confounding by wheeze, or similar.

Reviewer #2: The authors at John Radcliffe Hospital conducted this prospective observational study aiming to develop predictive models for exacerbation treatment outcome for patients with asthma and COPD exacerbation. They included 81 patients in their final analysis (59 asthma patients and 22 COPD patients). They first did a univariate analysis comparing the characteristics of patients who did or didn't receive systemic corticosteroids followed by multivariable logistic regression model to predict biological and historic predictors of physicians' decision to prescribe corticosteroids. After that they performed random forest models to find the predictors of treatment failure. I have three major concerns about this manuscript behind my recommendation to reject this manuscript:

Second, I find it methodologically troublesome to combine COPD and asthma patients in one basket and try to extrapolate prediction models from this combined group. While both asthma and COPD are obstructive lung diseases, there are big differences in the pathophysiology and patient populations. another minor point here, it's not clear how the diagnosis of COPD and asthma was made. did the patients have to have a pulmonary function test prior to the Ed visit or did they have to have radiologic evidence of emphysema for COPD at least?

Third, generating reliable prediction models requires larger sample size. As a clinician, if I read this article as a reader, I won't be able to draw reliable conclusions or use the prediction model presented in this manuscript as it was derived from 81 patients with no validation.

**********

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: No

Reviewer #2: No

PLoS One. 2021 Aug 20;16(8):e0254425. doi: 10.1371/journal.pone.0254425.r004

Author response to Decision Letter 1

24 Mar 2021

To the Editor, Dr Eugene Demidenko,

Thank you very much for the opportunity you have provided us to respond to the reviewers. We have taken the time to make the changes/additions and explanations suggested by the reviewers. The main changes are 1) we have now consistently used just one type of algorithm, the random forest, throughout the manuscript; 2) expanded upon limitations in our Discussion section; 3) expanded upon the relevance of our study in our Discussion section. We have uploaded a tracked changes version and an unmarked version both for the main manuscript and for the supplement.

Please see our responses to the two reviewers below:

Reviewer 1

Overall comments: “Overall, I don't think the methodology has been sufficiently thought through. To give you the benefit of the doubt, the wide mix of analyses and methods mean that you have not been able to thoroughly describe them. I think your aims are a little unclear - are you just looking for factors associated with SCS prescription and treatment failure, or are you trying to predict them? If the latter, the methodology is inappropriate because you are not reporting or testing the performance in a valid way, and the models are not being developed well. If the former, reporting the AIC and AUC seems unnecessary, as does comparing the results across different numbers of variables included. I think you need to refine your analysis question, and remove some of the analyses.”

Response: Thank you for your reflections and helpful comments. We have addressed them and we believe that the clarity of the work has been improved. This study is aimed at exploring: 1) variables associated with SCS prescription; 2) variables associated with treatment failure; 3) a multivariate model for predicting SCS prescription and predicting treatment failure. In keeping with your comments and for clarity, we have now used only one method for SCS prescription and treatment failure (random forest). We now discuss in the Methods and Discussion section the use of random forest models in our study and their strengths and limitations. We compare random forest models using different variable subset combinations in order to determine which models are both parsimonious and perform with high accuracy as per suggestion by the reviewer. We have removed all logistic regression analyses, including the corresponding AIC/ AUC which was requested by the previous reviewer.

Comment 1: “Abstract: SCS acronym used in abstract without introduction”.

Response 1: Thank you for spotting. This has now been addressed (page 3: lines 38-39) and used throughout after introduction.

Comment 2: “Introduction: please define the ‘standard initial treatment’”

Response 2: We have now added the clarification “as per NICE guidelines (with SCS and/or antibiotics)” with accompanying references on page 4, lines 62-63.

Comment 3: “Methods: Were there any participants with asthma-COPD overlap syndrome? This is not mentioned at all.”

Response 3: Thank you for your suggestion. We did not include this term since the asthma-COPD overlap syndrome term has been removed from any guidance and in UK clinical practice is not used.

Comment 4: “Methods: It is unclear to me why the methods were not consistent between the two main analyses: SCS prescription and treatment failure. You justify the use of the random forest for the second analysis, but it seems like it would have been appropriate for the first analysis for the same reason.”

Response 4: Thank you very much for your helpful comment. We fully agree with you that a random forest is the better model to use in both cases. Throughout all of the manuscript, random forest methodology and results reporting has now been used in place of a logistic regression analysis.

Comment 5: “Methods: Insufficient detail is provided on the construction of the multivariate logistic regression models. Did you try multiple combinations of between two and four variables and then select the best performing? Why was a stepwise approach not used?”

Response 5: As per comment 4, we have now employed a random forest model both for SCS prescription and treatment failure prediction. Details of the random forest methodology are provided in the Methods section of the main manuscript and in the S1 Appendix Statistical Analysis (main manuscript page 7, lines 129-130 and page 8, lines 133-135; supplement page 2, lines 19-51).

Comment 6: “Methods: By only allowing the random forest to choose from variable significant in the univariate logistic regression, potential interactions between variables are very limited. The ability to calculate such interactions is a key strength of tree-based algorithms.”

Response 6: Thank you. We agree that the random forest model with feature elimination subsequently enables the best combination of variables to be used (taking into account interactions between variables) for the final SCS prescription prediction and treatment failure prediction models.

Comment 7: “Methods: Which implementation of the random forest algorithm was used? What were the hyperparameters?”

Response 7: Thank you. We have added the implementation and model hyperparameter details in the S1 Appendix Statistical Analysis section (page 2, lines 20-23). The number of trees in the random forest models is 1000. The number of variables to be randomly selected at each split in each tree node was equal to the square root of the number of predictors being considered (number of predictors depend on the step of the feature elimination process).

Comment 8: “Methods: Did you look at the analyses stratified by diagnosis? It’s reasonable to assume it might differ between asthma and COPD.”

Response 8: We intentionally did not perform the predictive analyses stratified by diagnosis. This is for two reasons: 1) Out of the 81 patients, only 22 had a primary diagnosis of COPD. This would be too small a sample size for separate analysis; 2) whether the primary diagnosis was asthma or COPD was a variable considered in our SCS prescription and treatment failure prediction analysis. However, this variable was not selected in the best models, so it can be mathematically seen that this is not related to SCS prescription and treatment failure in our study cohort. This point is also discussed in the paper (page 20, lines 343-358).

Comment 9: “Results: Please change ‘asthmatics’ to ‘people with asthma’”

Response 9: Thank you – this has now been addressed (page 9, line 146).

Comment 10: “Results: No p-values appeared to be in bold in any of the tables.”

Response 10: Thank you – this has now been addressed for tables throughout the manuscript. No p-values are in bold and the table footnotes no longer mention that any p values are bolded.

Comment 11: “Results: I disagree with the previous reviewer that presenting the p-values in standard form is better, I think your previous approach was better.”

Response 11: Thank you – we also preferred the non-standard form. We have now reported p-values to 2dp but in keeping with PLOS ONE submission guidelines, p-values less than 0.001 are reported as p < 0.001. For tables where this applies, we have now indicated this in the table footnotes. We are happy to follow the Journal Editors selection as required.

Comment 12: “Results: Table 2 just says LAMA, but table 6 says lama use. Be consistent.”

Response 12: Thank you for spotting – we have now addressed this and used “LAMA” throughout all tables.

Comment 13: “Results: Table 2 – can you group the characteristics by type (continuous, binary etc) for ease of reading?”

Response 13: We have now addressed this.

Comment 14: “Results: Please define acronyms used in tables in the notes underneath them”

Response 14: Thank you - we have now addressed this throughout.

Comment 15: “Discussion: You should not be defining acronym for the first time in the discussions when it has been used throughout, such as PPI.”

Response 15: Thank you – we have now defined PPI in table footnotes in the Results section.

Comment 16: “Discussion: Grammatical error: “so the threshold to declare this, is reached more frequently”

Response 16: Thank you – we have deleted the comma (see page 19, line 312).

Comment 17: “Discussion: Overly strong assumption of causal relationship between eosinophilic inflammation and SCS prescribing. It seems very possible that this is confounding by wheeze, or similar.”

Response 17: We have now added that “…it is possible that increased wheeze at exacerbation is a confounder for the relationship between peripheral blood eosinophil percentage and SCS prescription.” (see page 17, lines 272-273). In addition, this is discussed in more detail on page 19 line 335 and page 20 lines 336-339.

Reviewer 2

Overall comments: “The authors at John Radcliffe Hospital conducted this prospective observational study aiming to develop predictive models for exacerbation treatment outcome for patients with asthma and COPD exacerbation. They included 81 patients in their final analysis (59 asthma patients and 22 COPD patients). They first did a univariate analysis comparing the characteristics of patients who did or didn't receive systemic corticosteroids followed by multivariable logistic regression model to predict biological and historic predictors of physicians' decision to prescribe corticosteroids. After that they performed random forest models to find the predictors of treatment failure. I have three major concerns about this manuscript behind my recommendation to reject this manuscript:

First, I don't understand what is the utility of finding the predictors whether or not to prescribe systemic corticosteroids for asthma and COPD exacerbation? The authors cited in the introduction that there is "a paucity of data which demonstrate inconsistent benefits for SCS and/or antibiotics for treating exacerbations of COPD and asthma in which treatment is received in hospital". The use of systemic steroids in asthma and COPD exacerbation is standard of care and has grade A evidence in the guidelines. There is grade A evidence (quote from GOLD guidelines) that "SCS use in COPD exacerbation can improve FEV1, oxygenation and shorten recovery time and hospitalization duration". As a pulmonologist, I believe if the diagnosis of asthma or COPD exacerbation is confirmed or highly suspected, the patient should be given SCS unless there is a contraindication or a reason for the treating physician not to give it. In this study, 21% of the patients didn't get corticosteroids. Additionally, 42% of the asthma exacerbation patients received antibiotics which isn't standard of care for asthma exacerbation patients which make me suspect there was some clinical suspicion for infection/pneumonia in these patients. Second, I find it methodologically troublesome to combine COPD and asthma patients in one basket and try to extrapolate prediction models from this combined group. While both asthma and COPD are obstructive lung diseases, there are big differences in the pathophysiology and patient populations. another minor point here, it's not clear how the diagnosis of COPD and asthma was made. did the patients have to have a pulmonary function test prior to the Ed visit or did they have to have radiologic evidence of emphysema for COPD at least? Third, generating reliable prediction models requires larger sample size. As a clinician, if I read this article as a reader, I won't be able to draw reliable conclusions or use the prediction model presented in this manuscript as it was derived from 81 patients with no validation.”

Response:

Thank you for your detailed reflections; we have responded below to each of the points.

1) Although SCS prescription is guideline recommended, it is clear that there is a real concern with regard to harm and a personalised approach to SCS is being evaluated (Bafadhel et al, AJRCCM 2011; Sivapalan LRM 2019). Furthermore, the evidence from Cochrane reviews heavily relies upon data from several decades and does not prevent further improvements being sought. In the real world not all patients will receive SCS at time or presentation with an exacerbation and this study is aimed at trying to determine the factors which can inform the decision to prescribe SCS and thus take a personalised individual patient approach. We have discussed these comments in the revised manuscript (page 4, lines 64-77).

2) We agree that it would appear that antibiotics prescription for asthma exacerbations are high; however, this is not an uncommon occurrence, as has previously been shown (Bafadhel et al, Chest 2011). We believe this is because patients are not seen by specialists in ED and reflects clinical practice. We have discussed this further in the manuscript and relate it to why a tool to help non-specialists is advantageous (page 18, lines 301-305).

3) In the UK, 15% of COPD is diagnosed on admission, whilst spirometry is not performed in over 40% of patients in the community. This would lead to a large proportion of patients presenting without spirometric classification and is reflective of clinical care. Combining this in this study reflects how an algorithm could help non-specialists. This has been discussed in the manuscript (page 20, lines 343-348).

4) Diagnosis was clinical history recorded from the patient and their medical records if available.

5) Regarding your point as to the lack of validation, in the main manuscript Discussion section we have clarified that performance of our random forest algorithms in our study was reported on test patients (see page 20, 357-359 and page 21 line 360). Performance of the models is tested on a validation set as per the supplement (page 2, lines 28-51), but we agree should be further validated in an external validation set, as per our discussion (main manuscript page 20, line 356).

Thank you again for your consideration of our manuscript.

We look forward to hearing from you.

Kind regards,

Mr Andreas Halner, Dr Richard Russell and Professor Mona Bafadhel

on behalf of the co-authors

PLoS One. doi: 10.1371/journal.pone.0254425.r005

Decision Letter 2

Eugene Demidenko

28 Jun 2021

Predicting Treatment Outcomes Following an Exacerbation of Airways Disease

PONE-D-20-36305R2

Dear Dr. Halner,

Both reviewers indicated "All comments have been addressed" although the first reviewer suggested "Rejection." I feel that you indeed addressed all comments and critique and therefore made the decision to accept the paper. Congratulations!

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,

Eugene Demidenko, Ph.D.

Academic Editor

PLOS ONE

Additional Editor Comments (optional):

Reviewers' comments:

Reviewer's Responses to Questions

Comments to the Author

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?

Reviewer #1: Yes

Reviewer #2: (No Response)

**********

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

Reviewer #1: Yes

Reviewer #2: (No Response)

**********

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

Reviewer #1: No

Reviewer #2: (No Response)

**********

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

Reviewer #1: Yes

Reviewer #2: (No Response)

**********

6. Review Comments to the Author

Reviewer #1: 1. please capitalise 'Shapiro-Wilk' and 'Wilkcoxon; (which should replace wilcox)

2. I still think that you have not justified the analysis of the factors associated with SCS prescription. Some of these factors are likely to be confounding, and I still think that a qualitative survey of health care professionals would be better placed to answer this question. I would strongly suggest removing this component of the paper unless you are able to better justify it.

3. Similarly, I don't think you have sold me on the value of this analysis. I see that PPI as a protective factor for treatment failure is a nice finding, but you are not the first to report this - see this for example: https://jamanetwork.com/journals/jamainternalmedicine/article-abstract/227086. I need to see some discussion about the clinical implications of this research. What are you proposing should be done as a result of your study? If you are highlighting people who are at a higher risk of treatment failure (although you have not provided a decision tool to predict this either, just reported some risk factors) then what is your alternative proposal?

Reviewer #2: I salute the authors for their responses and congratulate them for this work. I feel they have provided adequate responses to my previous concerns.

**********

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: No

Reviewer #2: No

PLoS One. doi: 10.1371/journal.pone.0254425.r006

Acceptance letter

Eugene Demidenko

5 Aug 2021

PONE-D-20-36305R2

Predicting Treatment Outcomes Following an Exacerbation of Airways Disease

Dear Dr. Halner:

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. Eugene Demidenko

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 Appendix. Predicting treatment outcomes following an exacerbation of airways disease appendix.

(DOCX)

Click here for additional data file.^{(65KB, docx)}

S1 Dataset

(CSV)

Click here for additional data file.^{(25.6KB, csv)}

Data Availability Statement

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

[pone.0254425.ref001] 1.Vogelmeier CF, Criner GJ, Martinez FJ, et al. Global strategy for the diagnosis, management, and prevention of chronic obstructive lung disease 2017 report. Am J Respir Crit Care Med. 2017; 195: 557–582. doi: 10.1164/rccm.201701-0218PP [DOI] [PubMed] [Google Scholar]

[pone.0254425.ref002] 2.Chronic obstructive pulmonary disease in over 16s: diagnosis and management | Guidance | NICE. (2010). [PubMed]

[pone.0254425.ref003] 3.Greening NJ, Williams JEA, Hussain SF, et al. An early rehabilitation intervention to enhance recovery during hospital admission for an exacerbation of chronic respiratory disease: Randomised controlled trial. BMJ. 2014; 349(7967): G4315. doi: 10.1136/bmj.g4315 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pone.0254425.ref004] 4.Scenario: Acute exacerbation of asthma. 2020. Available from: https://cks.nice.org.uk/topics/asthma/management/acute-exacerbation-of-asthma/.

[pone.0254425.ref005] 5.Walters JAE, Tan DJ, White CJ, Gibson PG, Wood-Baker R, Walters EH. Systemic corticosteroids for acute exacerbations of chronic obstructive pulmonary disease. Cochrane Database Syst Rev. 2014; (9):CD001288. doi: 10.1002/14651858.CD001288.pub4 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pone.0254425.ref006] 6.Vollenweider DJ, Frei A, Steurer-Stey CA, Garcia-Aymerich J, Puhan MA. Antibiotics for exacerbations of chronic obstructive pulmonary disease. Cochrane Database Syst Rev. 2018; 10(10):CD010257. doi: 10.1002/14651858.CD010257.pub2 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pone.0254425.ref007] 7.Phipatanakul W, Mauger DT, Sorkness RL, et al. Effects of Age and Disease Severity on Systemic Corticosteroid Responses in Asthma. Am J Respir Crit Care Med. 2017; 95(11):1439–1448. doi: 10.1164/rccm.201607-1453OC [DOI] [PMC free article] [PubMed] [Google Scholar]

[pone.0254425.ref008] 8.Normansell R, Sayer B, Waterson S, Dennett EJ, Del Forno M, Dunleavy A. Antibiotics for exacerbations of asthma (Review). Cochrane Database Syst Rev. 2018; (6): CD002741. doi: 10.1002/14651858.CD002741.pub2 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pone.0254425.ref009] 9.Waljee A, Rogers M, Lin P, et al. Short term use of oral corticosteroids and related harms among adults in the United States: population based cohort study. BMJ 2017;357:j1415. doi: 10.1136/bmj.j1415 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pone.0254425.ref010] 10.Bafadhel M, McKenna S, Terry S, et al. Blood eosinophils to direct corticosteroid treatment of exacerbations of chronic obstructive pulmonary disease: A randomized placebo-controlled trial. Am. J. Respir. Crit. Care Med. 2012; 186(1): 48–55. doi: 10.1164/rccm.201108-1553OC [DOI] [PMC free article] [PubMed] [Google Scholar]

[pone.0254425.ref011] 11.Sivapalan P, Lapperre TS, Janner J, et al. Eosinophil-guided corticosteroid therapy in patients admitted to hospital with COPD exacerbation (CORTICO-COP): a multicentre, randomised, controlled, open-label, non-inferiority trial. Lancet Respir. Med. 2019; 7(8), 699–709. doi: 10.1016/S2213-2600(19)30176-6 [DOI] [PubMed] [Google Scholar]

[pone.0254425.ref012] 12.Vestbo J, Hurd SS, Agusti AG, et al. Global strategy for the diagnosis, management, and prevention of chronic obstructive pulmonary disease GOLD executive summary. Am J Respir Crit Care Med. 2013; 187:347–365. doi: 10.1164/rccm.201204-0596PP [DOI] [PubMed] [Google Scholar]

[pone.0254425.ref013] 13.British guideline on the management of asthma. (2014). [Google Scholar]

[pone.0254425.ref014] 14.Brightling CE, Monterio W, Green RH, et al. Induced sputum and other outcome measures in chronic obstructive pulmonary disease: Safety and repeatability. Respir. Med. 2001; 95(12): 999–1002. doi: 10.1053/rmed.2001.1195 [DOI] [PubMed] [Google Scholar]

[pone.0254425.ref015] 15.Fletcher CM, Clifton M, Fairbairn AS, et al. Standardised questionnaire on respiratory symptoms: a statement prepared and approved by the MRC Committee on the Aetiology of Chronic Bronchitis (MRC breathlessness score). BMJ. 1960; 2(5213): 1665. 13688719 [Google Scholar]

[pone.0254425.ref016] 16.Zigmond AS, Snaith RP. The Hospital Anxiety and Depression Scale. Acta Psychiatr. Scand. 1983; 67(6): 361–370. doi: 10.1111/j.1600-0447.1983.tb09716.x [DOI] [PubMed] [Google Scholar]

[pone.0254425.ref017] 17.Group EuroQol. EuroQol—a new facility for the measurement of health-related quality of life. Health Policy. 1990; 16(3): 199–208. doi: 10.1016/0168-8510(90)90421-9 [DOI] [PubMed] [Google Scholar]

[pone.0254425.ref018] 18.R Core Team (2019). R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. URL https://www.R-project.org/. [Google Scholar]

[pone.0254425.ref019] 19.Breiman L. Random forests. Mach. Learn. 2001; 45(1): 5–32. [Google Scholar]

[pone.0254425.ref020] 20.Kuhn M, Johnson K. Applied Predictive Modeling. New York: Springer; 2013. [Google Scholar]

[pone.0254425.ref021] 21.Ding B, Small M, Bergström G, Holmgren U. COPD Symptom Burden: Impact on Health Care Resource Utilization, and Work and Activity Impairment. Int J Chron Obstruct Pulmon Dis. 2017; 12: 677–89. doi: 10.2147/COPD.S123896 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pone.0254425.ref022] 22.Hurst JR, Vestbo J, Anzueto A, et al. Susceptibility to Exacerbation in Chronic Obstructive Pulmonary Disease. N Engl J Med. 2010; 363: 1128–138. doi: 10.1056/NEJMoa0909883 [DOI] [PubMed] [Google Scholar]

[pone.0254425.ref023] 23.Sakae TM, Pizzichini MMM, Teixeira PJZ, Maurici Da Silva R, Trevisol DJ, Pizzichini E. Exacerbations of COPD and Symptoms of Gastroesophageal Reflux: A Systematic Review and Meta-analysis. J Bras Pneumol. 2013; 39(3): 259–71. doi: 10.1590/S1806-37132013000300002 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pone.0254425.ref024] 24.Nishimura K, Izumi T, Tsukino M, Oga T. Dyspnea is a better predictor of 5-year survival than airway obstruction in patients with COPD. Chest. 2002; 121(5): 1434–1440. doi: 10.1378/chest.121.5.1434 [DOI] [PubMed] [Google Scholar]

[pone.0254425.ref025] 25.Steer J, Norman EM, Afolabi OA, Gibson GJ, Bourke SC. Dyspnoea severity and pneumonia as predictors of in-hospital mortality and early readmission in acute exacerbations of COPD. Thorax. 2012; 67(2): 117–21. doi: 10.1136/thoraxjnl-2011-200332 [DOI] [PubMed] [Google Scholar]

[pone.0254425.ref026] 26.Mantero M, Rogliani P, Di Pasquale M, et al. Acute exacerbations of COPD: Risk factors for failure and relapse. Int J Chron Obstruct Pulmon Dis. 2017; 12: 2687–2693. doi: 10.2147/COPD.S145253 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pone.0254425.ref027] 27.Chronic obstructive pulmonary disease (acute exacerbation): antimicrobial prescribing (NG114) (2018).

[pone.0254425.ref028] 28.Bafadhel M, Clark T, Reid C, et al. Procalcitonin and C-reactive protein in hospitalized adult patients with community-acquired pneumonia or exacerbation of asthma or COPD. Chest. 2010; 139(6):1410–1418. doi: 10.1378/chest.10-1747 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pone.0254425.ref029] 29.Prins HJ, Duijkers R, van der Valk P, et al. CRP-guided Antibiotic Treatment in acute exacerbations of COPD admitted to Hospital. Eur Respir J. 2019; 53(5):180201. [DOI] [PubMed] [Google Scholar]

[pone.0254425.ref030] 30.Boersma WG. Antibiotics in acute exacerbations of COPD: The good, the bad and the ugly. Eur Respir J. 2012; 40(1): 1–3. doi: 10.1183/09031936.00211911 [DOI] [PubMed] [Google Scholar]

[pone.0254425.ref031] 31.Trivedi SG, Lloyd CM. Eosinophils in the Pathogenesis of Allergic Airways Disease. Cell Mol Life Sci. 2007; 64(10): 1269–289. doi: 10.1007/s00018-007-6527-y [DOI] [PMC free article] [PubMed] [Google Scholar]

[pone.0254425.ref032] 32.Houston JC, De Navasquez S, Trounce JR. A Clinical and Pathological Study of Fatal Cases of Status Asthmaticus. Thorax. 1953; 8:(3): 207–13. doi: 10.1136/thx.8.3.207 [DOI] [PMC free article] [PubMed] [Google Scholar]

PERMALINK

Predicting treatment outcomes following an exacerbation of airways disease

Andreas Halner

Sally Beer

Richard Pullinger

Mona Bafadhel

Richard E K Russell

Roles

Abstract

Background

Aims

Methods

Results

Conclusion

Introduction

Methods

Participants

Study design

Measurements

Statistical analyses

Results

Table 1. Characteristics of the asthma and COPD participants at exacerbation.

Characteristics according to D0 treatment

Table 2. Comparison of characteristics of participants given SCS only versus participants given antibiotics only versus participants given both SCS and antibiotics versus participants given neither SCS nor antibiotics.

Predicting physician SCS prescription decision

Table 3. Comparison of D0 exacerbation characteristics of patients given SCS versus patients not given SCS for which p < 0.1 (see S1 Appendix for all other characteristics evaluated).

Predicting a treatment failure

Table 4. Comparison of D0 exacerbation characteristics of patients who failed treatment with those who succeeded treatment which p < 0.1 (see S1 Appendix for all other characteristics evaluated).

Table 5. Random forest importance scores of variables in multivariate treatment failure prediction models for all participants.

Discussion

Conclusion

Supporting information

Acknowledgments

Data Availability

Funding Statement

References

Decision Letter 0

Eugene Demidenko

Roles

Author response to Decision Letter 0

Decision Letter 1

Eugene Demidenko

Roles

Author response to Decision Letter 1

Decision Letter 2

Eugene Demidenko

Roles

Acceptance letter

Eugene Demidenko

Roles

Associated Data

Supplementary Materials

Data Availability Statement

ACTIONS

PERMALINK

RESOURCES

Similar articles

Cited by other articles

Links to NCBI Databases