The study was conducted on data from individuals recovering from an ECOPD, either cared for in acute hospitals and transferred to Rehabilitation Hospitals or cared for at home by their general practitioners [GP], and admitted after an out-patient clinic visit, between July, 1st, 2018 to December, 31st, 2021 to hospitals of the network previously described.15 These hospitals share common indications, evaluation, diagnostic and management tools and protocols for PR. During the pandemic (March 2020–December 2021), only participants with negative swab tests were admitted. It's worth noting that this study utilized the same dataset analyzed and published elsewhere.15

The study included data from individuals with:

• -

  Diagnosis of COPD according to the GOLD guidelines.2

• -

  Persistent breathlessness and/or exercise limitation within the previous 30 days after an ECOPD managed in acute care hospitals or within the previous 4 weeks after an ECOPD managed at home by the GP.16

• -

  Present stable conditions, defined as absence of acute worsening in symptoms (i.e. no change in dyspnea, cough, and/or sputum beyond the day-to-day variability) that would have required a change in management compared to the conditions reported at home or at discharge from the referring acute care hospital.16

• -

  Availability of data on lung function and paired pre and post-PR data of outcome measures.

• -

  Participating in at least 12 PR sessions, a threshold chosen according to our previous published experience.17

Exclusion criteria from PR had been: severe comorbidities such as oncological, neurological disorders, heart failure, or recent (less than 4 months) acute ischemic cardiovascular diseases with an unstable status and individuals unable or refusing to perform PR.

The following data were recorded at admission: demographics, anthropometrics, history of ECOPD in previous 12 months, Comorbidity Index of the Cumulative Illness Rating Scale (CIRS), BMI – airflow obstruction – dyspnea, and exercise capacity (BODE) index, provenience (hospital or home), length of rehabilitation hospital stay (LoS), occurrence of chronic respiratory failure (CRF), distribution in GOLD stages, drug therapy in stable state.

Before the program the following assessments had been performed:

• -

  Forced expiratory volumes according to standards.18

• -

  Functional disability by the Barthel index (BI).19

Before and after the program, the following outcome measures had been assessed:

• -

  Six-minute walking distance test (6MWT).20,21 The minimal clinically important difference (MCID) has been reported as an improvement by at least 30m.20

• -

  Dyspnea by the Medical Research Council (MRC) scale.22 A one-point reduction is considered equivalent to MCID.23

• -

  Dyspnea by the Barthel Index dyspnea (BId). The MCID has been defined as a 9-point reduction for individuals without and as a 12-point reduction for individuals with CRF, respectively.24

• -

  COPD assessment test (CAT). A two-point reduction in score has been reported as the MCID.25

Our program is supervised by multidisciplinary teams of trained and experienced chest physicians, nurses, physical therapists, dieticians, and psychologists full-time dedicated to PR. It starts within 2 days from admission, after baseline evaluations and includes daily supervised sessions (6 weekly days) of cycle training according to Maltais et al.,26 until performing 30min of continuous cycling at 50–70% of maximal load, (calculated on baseline 6MWT according to Luxton et al.27). Workload was increased by 5 Watts when subjects scored their dyspnea or leg fatigue as <3 on a 10-point Borg Scale, remained unchanged if score was 4 or 5 and was reduced for scores of >5. Pulse oximetry, arterial blood pressure, and heart rate are monitored during sessions.

The program also includes optimization of medications, education, nutritional programs, and psychosocial counseling when appropriate, abdominal, upper, and lower limb muscle activities lifting weights progressively. The duration of daily activities is 2–3h. The program is performed in Gymn Room with full availability of safety tools (e.g. CPR).

During pandemic, protective measures were adopted, such as personal protective equipment, distance among individuals not less than 2m, disinfection of materials, frequent air changes, execution of a swab at first harmful signs.

Quantitative variable distribution is described as median (25th, 75th percentiles) since most of them deviated from the normality assumptions, as determined by visual inspection of histograms. Categorical variable distribution is described as absolute and relative (%) frequencies. There were no missing values in the data. The whole dataset was randomly split into a training (70% of the whole dataset) and a test set (30%). Statistical and ML methods for regression (linear regression, quantile regression [quantile=0.5], regression trees, and conditional inference trees) were trained to predict quantitative pre to post PR changes in outcomes and tested as described in the Supplementary material – Supplementary Methods section. Post PR changes in outcomes were computed as value at discharge – value at admission and were dichotomized based on the corresponding MCID. Predictive performances were expressed as mean absolute error (MAE), sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV) (Supplementary material – Supplementary Methods). Significance level was set to α=0.05. Statistical and ML analyses were performed by the R software tool version 4.2.2 (www.r-project.org).

Characteristics of 1159 individuals included are in Table 1.15 The majority of participants were males, were admitted from home, used triple (Long acting muscarinic antagonists+Long acting beta agonists+inhaled corticosteroids: LAMA+LABA+ICS) inhaled therapy and were included in most severe GOLD stages (1 and 2 or C and D) but did not suffer from CRF.

Performances of ML methods in predicting post-PR changes in outcome measures using values at admission of explanatory variables (reported in Supplementary Table 1) were assessed and compared on training set data (n=811) as described in Supplementary material. The evaluated approaches showed similar median prediction error estimates (Supplementary material – Supplementary Fig. 1).

Quantile regression coupled with a backward features selection approach based on statistical significance (QR-p) was selected as the most suitable method to predict changes (Supplementary material – Supplementary Fig. 2).

Regression coefficients estimated by the four outcome-specific multivariate regressions on the whole training set using post-PR changes in outcomes are reported in Table 2.

Multivariate regression allowed estimating a post-PR 0.17m reduction in median 6MWT change for every meter of 6MWT at admission (p-value<0.0001). Similarly, for every year increase in age and every point in admission BId, we expected a reduction in median post PR 6MWT change by −0.71 and −0.35m respectively (p<0.05). On the opposite, for each unit FEV1/FVC, % a 0.34-m increase in post-PR median 6MWT change was predicted (p=0.0191). Also, individuals from hospitals showed a 24.85-m increase in median post-PR change in 6MWT compared to individuals from home (p-value<0.0001).

The corresponding equation to be used to predict 6MWT post PR change is the following: 6MWT post PR change (meters)=135.124−0.171×6MWT at admission (m)+24.849×Provenience (Hospital=1, Home=0)−0.707×Age (years)−0.351×BId at admission (points)+0.339×FEV1/FVC %.

Taken together, these results indicate that the higher is the baseline exercise limitation, the higher is the improvement. Aged as well as dyspnoeic individuals share a lower predicted improvement, while individuals with less severe airway obstruction as well as those admitted from hospitals have larger potential improvements in exercise capacity. Multivariate explainability analyses identified 6MWT at admission (meters) as the strongest predictor of post-PR changes in 6MWT, followed by provenience from the hospital, age, BId, and FEV1/FVC, % at admission.

Each point of CAT and MRC and each meter of 6MWT at admission correspond to a decrease in the post-program median change in CAT by −0.41, −0.47, and −0.01 points respectively (p<0.05, Table 2).

The equation is the following: CAT post PR change (points)=5.263−0.407 * CAT at admission (points)−0.469*MRC at admission (points)−0.009*6MWT at admission (m).

Thus, individuals with more disease impact, more dyspnea, and more exercise limitation have a greater potential improvement in CAT. Multivariate explainability analyses revealed that CAT at admission was the strongest predictor of post-PR changes in CAT, followed by 6MWT and MRC at admission.

Similarly, each point increase in BId and MRC at admission corresponds to a post-PR decrease by −0.19 and −1 points in median BId and MRC, respectively (p-value<0.0001, Table 2).

The corresponding equations are the following:

• a)

  BId post PR change (points)=−4.038−0.192*BId at admission (points)

• b)

  MRC post PR change (points)=2−1*MRC at admission (points)

Individuals with more severe dyspnea during Activity daily life (ADL) at admission are more prone to improve.

The ability of the four regression models to correctly predict post-PR changes in outcomes was then assessed on the independent test set data (n=348). Mean absolute error estimates of the four models were 44.70m, 3.22 points, 5.35 points, and 0.32 points for 6MWT, CAT, BId and MRC.

Fig. 1 shows remarkable discrepancies between predicted and observed post PR changes in outcomes, especially for low and high observed outcome distribution values, particularly for 6MWT and BId. In detail, in comparison to the observed post-PR changes in 6MWT ranging between −243 and +348m, the spectrum of predicted values of change in 6MWT is narrower, ranging between −8 and +139m. Similarly, observed changes in BId ranged between −68 and +11 points, whereas predicted changes ranged between −21 and −4 points. Post-PR changes in CAT ranged between −27 and +4 points, as compared to predicted values between −15 and +1 points. Predicted post-PR changes in MRC ranged between −3 and +1 points, whereas observed changes ranged between −2 and +2 points.

Fig. 1.

Comparison between observed and predicted outcome variables change on the test set. Legend. Black dots represent pointwise predictions while gray lines represent the corresponding 95% confidence intervals; diagonal dashed lines represent the theoretical condition of perfect agreement between observed and predicted values. Abbreviations. 6MWT: six-minute walking distance test; CAT: COPD assessment test; BId: Barthel Index dyspnea; MRC: Medical Research Council; m: meters.

(0.36MB).

Fig. 2 shows MAE and outcome changes by five intervals characterized by approximately equal frequency of participants (quintiles) based on outcome variables distribution at admission. Prediction error tends to be higher for participants with lower 6MWT, as well as higher CAT and BId values at admission, with these subsets of participants characterized by the largest median post-PR changes.

Fig. 2.

Prediction error and outcome change from admission by intervals of outcome variables at admission. Legend. The three lines plots in panels A–D describe graphically the Mean Absolute Error (MAE) by quintile of outcome variables distribution at admission (6MWT, CAT, and BId) or by MRC value characterizing the evaluated machine learning methods. MRC changes have been not analyzed by quintiles of MRC at admission but by pooling values at admission between 0 and 2 to define approximately balanced intervals. The three boxplots in panels E–H describe the outcomes’ change between admission and discharge distribution by quintile of outcome variables distribution at admission. Each boxplot describes from bottom to top: the lowest non-outlier value/minimum value; 25th percentile; median value; 75th percentile and the highest non-outlier value/maximum value; outliers are depicted as dots. The black continuous horizontal line corresponds to the condition of no change between admission and discharge; the black dashed horizontal line corresponds to the MCID value for each outcome variable and gray-shaded areas correspond to the condition of MCID achievement. Abbreviations. 6MWT: six-minute walking distance test; CAT: COPD assessment test; BId: Barthel Index dyspnea; MRC: Medical Research Council; m: meters; CRF: chronic respiratory failure.

(0.32MB).

In detail, MAE characterizing quantile regression predicting 6MWT at discharge was 67.59m in individuals with admission 6MWT<173m (median 6MWT change=+92.5m) whereas 30.47m in those with admission 6MWT between 423 and 600m (median 6MWT change=+25.5m) (Fig. 2A and E).

When predicting CAT at discharge, MAE was 2.11 points in individuals with admission CAT<11 (median CAT change=−2.5 points), increasing up to 4.98 points in those with CAT ranging from 25 to 35 (median CAT change=−10 points) (Fig. 2B and F).

Similarly, MAE was 3.01 points in people with admission BId<13 (median BId change=−3 points), increasing up to 10.35 points in those with admission BId ranging from 43 to 89 (median BId change=−12 points) (Fig. 2C and G).

No interpretable trend in MAE was observed when analyzing predictions for MRC (Fig. 2D and H).

Post PR changes in outcomes have been discretized into binary values, indicating whether a participant reached the MCID or not. Performances in discriminating participants reaching from those not reaching MCID for the four outcomes are depicted in Table 3 and show sensitivity ranging from 61.78% to 98.99%, PPV ranging from 63.82% to 93.99%, specificity ranging from 14.00% to 71.20% and NPV ranging from 52.31% to 70.00%. The ability of quantitative predictions of post-PR changes to correctly identify participants reaching the MCID for the combination of outcomes was also assessed (Table 3).

Sensitivity, PPV, specificity and NPV in predicting the combined 6MWT-CAT-BId outcomes were 47.78%, 59.80%, 75.58%, and 80.58% respectively, with a 25.86% proportion of participants reaching MCID for 6MWT, CAT, and BId.

Sensitivity, PPV, specificity, and NPV in predicting the combined 6MWT-CAT-MRC outcome were 73.94%, 59.8%, 55.19% and 70.14% respectively, with a 47.41% proportion of participants reaching MCID for 6MWT, CAT, and MRC.

These results suggest a better capacity to accurately discriminate between individuals reaching or not the MCID of MRC as compared to BId. Consequently, there is a higher probability of correctly detecting individuals reaching the MCID for the combined 6MWT-CAT-MRC concerning the 6MWT-CAT-BId outcome.

The prediction models of different issues in PR have been addressed also in a few prospective studies.37,38 The main limitations of our study are related to its retrospective design. However, in addition to the relatively new approach of analysis, our study represents a real-life condition, and its results are supported by the large sample size at a time when even randomized controlled trials are being questioned.39

A control population not performing PR would have clarified whether any improvement would have been (also) time-dependent. However, given recognized benefits and mission of our hospitals, not performing PR would have been unaethical.