key: cord-0704720-yw66q4e7
authors: Sonnweber, Thomas; Tymoszuk, Piotr; Sahanic, Sabina; Boehm, Anna; Pizzini, Alex; Luger, Anna; Schwabl, Christoph; Nairz, Manfred; Grubwieser, Philipp; Kurz, Katharina; Koppelstätter, Sabine; Aichner, Magdalena; Puchner, Bernhard; Egger, Alexander; Hoermann, Gregor; Wöll, Ewald; Weiss, Günter; Widmann, Gerlig; Tancevski, Ivan; Löffler-Ragg, Judith
title: Investigating phenotypes of pulmonary COVID-19 recovery: A longitudinal observational prospective multicenter trial
date: 2022-02-08
journal: eLife
DOI: 10.7554/elife.72500
sha: 2fe7eb666cac19aba2482c5c9091a0e3fa28eb0e
doc_id: 704720
cord_uid: yw66q4e7

BACKGROUND: The optimal procedures to prevent, identify, monitor, and treat long-term pulmonary sequelae of COVID-19 are elusive. Here, we characterized the kinetics of respiratory and symptom recovery following COVID-19. METHODS: We conducted a longitudinal, multicenter observational study in ambulatory and hospitalized COVID-19 patients recruited in early 2020 (n = 145). Pulmonary computed tomography (CT) and lung function (LF) readouts, symptom prevalence, and clinical and laboratory parameters were collected during acute COVID-19 and at 60, 100, and 180 days follow-up visits. Recovery kinetics and risk factors were investigated by logistic regression. Classification of clinical features and participants was accomplished by unsupervised and semi-supervised multiparameter clustering and machine learning. RESULTS: At the 6-month follow-up, 49% of participants reported persistent symptoms. The frequency of structural lung CT abnormalities ranged from 18% in the mild outpatient cases to 76% in the intensive care unit (ICU) convalescents. Prevalence of impaired LF ranged from 14% in the mild outpatient cases to 50% in the ICU survivors. Incomplete radiological lung recovery was associated with increased anti-S1/S2 antibody titer, IL-6, and CRP levels at the early follow-up. We demonstrated that the risk of perturbed pulmonary recovery could be robustly estimated at early follow-up by clustering and machine learning classifiers employing solely non-CT and non-LF parameters. CONCLUSIONS: The severity of acute COVID-19 and protracted systemic inflammation is strongly linked to persistent structural and functional lung abnormality. Automated screening of multiparameter health record data may assist in the prediction of incomplete pulmonary recovery and optimize COVID-19 follow-up management. FUNDING: The State of Tyrol (GZ 71934), Boehringer Ingelheim/Investigator initiated study (IIS 1199-0424). CLINICAL TRIAL NUMBER: ClinicalTrials.gov: NCT04416100

The ongoing COVID-19 pandemic challenges health-care systems. As of December 2021, the John Hopkins dashboard (Dong et al., 2020) reports 276 million cases and 5.4 million COVID-19-related deaths worldwide (Johns Hopkins Coronavirus Resource Center, 2021) . Although the vast majority of COVID-19 patients display mild disease, approximately 10-15% of cases progress to a severe condition and approximately 5% suffer from critical illness (Perez-Saez, 2021; Huang et al., 2020) . Similar to severe acute respiratory syndrome (SARS) (Hui et al., 2005; Ng et al., 2004; Ngai et al., 2010; Lam et al., 2009) , a significant portion of COVID-19 patients report lingering or recurring clinical impairment and cardiopulmonary recovery may take several months to years Sahanic et al., 2021; Caruso et al., 2021; Huang et al., 2021b; Huang et al., 2021a; Faverio et al., 2021; Hellemons et al., 2021; Zhou et al., 2021; Venkatesan, 2021) . This observation has led to the introduction of the term 'long COVID,' defined by the persistence of COVID-19 symptoms for more than 4 weeks, and the 'post-acute sequelae of COVID-19' (PASC) referring to symptom persistence for more than 12 weeks Shah et al., 2021; Sudre et al., 2021b) . Evidence-based strategies for prediction, monitoring, and treatment of PASC are urgently needed (Raghu and Wilson, 2020) .

We herein prospectively analyzed the prevalence of nonresolving structural and functional lung abnormalities and persistent COVID-19-related symptoms 6 months after diagnosis. Using univariate risk modeling as well as multiparameter clustering and machine learning (ML), we investigated sets of risk factors and tested the operability of ML classifiers at predicting protracted lung and symptom recovery. The classification and prediction procedures were implemented in an opensource risk assessment tool (https://im2-ibk. shinyapps.io/CovILD/).

The CovILD ('Development of interstitial lung disease in COVID-19') multicenter, longitudinal observational study was initiated in April 2020. Adult residents of Tyrol, Austria, with symptomatic, PCR-confirmed SARS-CoV-2 infection (WHO, 2021) were enrolled by the Department of Internal Medicine II at the Medical University of Innsbruck (primary follow-up center), St. Vinzenz Hospital in Zams, and the acute rehabilitation facility in Münster ( Table 1) . The participants were diagnosed with COVID-19 between 3 March and 29 June 2020. In course of the study, including the 2020 SARS-CoV-2 outbreak and follow-up visits, the regional health system was able to guarantee an unrestricted, optimal standard of diagnostics and care for all participants. Corticosteroids were not standard of care during the recruitment period of the study, thus were not administered as a therapy of acute COVID-19. Some participants with nonresolving pneumonia received systemic steroids beginning from week 4 post diagnosis at the discretion of the physician ( Table 2 ). The analysis endpoints were the presence of any, mild (severity score ≤ 5), and moderate-to-severe (severity score > 5) lung computed tomography (CT) abnormalities, impaired lung function (LF), and persistent COVID-19 symptoms at the 180-day follow-up visit ( Table 3 ).

In total, 190 COVID-19 patients were screened for participation. Thereof, n = 18 subjects refused to give informed consent, n = 27 declared difficulties to appear at the study follow-ups. Data of n = 145 participants were eligible for analysis ( Figure 1 ). All participants gave written informed consent. The study was approved by the Institutional Review Board at the Medical University of Innsbruck (approval number: 1103/2020) and registered at ClinicalTrials. gov (NCT04416100).

We retrospectively assessed patient characteristics during acute COVID-19 and performed follow-up investigations at 60 days (63 ± 23 days [mean ± SD]; visit 1), 100 days (103 ± 21 days; visit 2), and 180 days (190 ± 15 days; visit 3) after diagnosis of COVID-19. Each visit included symptom and physical performance assessment with a standardized questionnaire, LF testing, standard laboratory testing, and a CT scan of the chest. The variables available for analysis with their stratification schemes are listed in Appendix 1-table 1.

Serological markers were determined in certified laboratories (Central Institute of Clinical and Chemical Laboratory Diagnostics, Rheumatology and Infectious Diseases Laboratory, both at the University Hospital of Innsbruck). C-reactive protein (CRP), interleukin-6 (IL-6), N-terminal pro natriuretic peptide (NT-proBNP), and serum ferritin were measured using a Roche Cobas 8000 analyzer. D-dimer was determined with a Siemens BCS-XP instrument using the Siemens D-Dimer Innovance reagent. Anti-S1/S2 protein SARS-CoV-2 immunoglobulin gamma (IgG) were quantified with LIAISON Table 2 . Hospitalization and medication during acute Outpatient (n = 36) Hospitalized (n = 37)

Hospitalized oxygen therapy (n = 40)

Hospitalized intensive care unit (n = 32)

Mean hospitalization time, days 0 (SD = 0) 6.9 (SD = 3.6) 11.8 (SD = 6.3) 34.8 (SD = 15.7)

Hospitalized >7 days 0% (n = 0) 43.2% (n = 16) 80% (n = 32) 100% (n = 32)

Anti-infectives 11.1% (n = 4) 45.9% (n = 17) 72.5% (n = 29) 87.5% (n = 28) Antiplatelet drugs 2.8% (n = 1) 10.8% (n = 4) 22.5% (n = 9) 25% (n = 8) Anticoagulatives 2.8% (n = 1) 2.7% (n = 1) 5% (n = 2) 15.6% (n = 5) Corticosteroids* † 2.8% (n = 1) 5.4% (n = 2) 22.5% (n = 9) 40.6% (n = 13) Immunosuppression ‡ † 0% (n = 0) 2.7% (n = 1) 5% (n = 2) 9.4% (n = 3) * From the week 4 post diagnosis on, at the discretion of the physician. † Subsumed under 'immunosuppression, acute COVID-19' for data analysis. ‡ Immunosuppressive medication prior to COVID-19. 

Statistical analyses were performed with R version 4.0.5 ( Figure 1) . Data transformation and visualization were accomplished by tidyverse (Wickham et al., 2019) , ggplot2 (Wickham, 2016) , ggvenn, plotROC (Sachs, 2017) , and cowplot (Wilke, 2019) packages. The recorded variables were binarized as shown in Appendix 1-table 1. Acute COVID-19 severity strata were defined as presented in Table 1 . p-Values were corrected for multiple comparisons with the Benjamini-Hochberg method (Benjamini and Hochberg, 1995) , and effects were termed significant for p<0.05.

Overlap between the 180-day follow-up outcome features was assessed by analysis of quasiproportional Venn plots (package nVennR) (Pérez-Silva et al., 2018) and calculation of the Cohen's κ statistic (package vcd) (Fleiss et al., 1969) . Kinetics of binary outcome variables in participants subsets with the complete longitudinal data record was modeled with mixed-effect logistic regression (random effect: individual, fixed effect: time, packages lme4 [Bates et al., 2015] and lmerTest [Kuznetsova et al., 2017] ). Analyses in the severity groups were done with separate models. Significance was assessed by the likelihood ratio test (LRT) against the random-term-only model. Univariate risk modeling was performed with fixed-effect logistic regression (Appendix 1-table 2). Odds ratio (OR) significance was determined by Wald Z test. In-house-developed linear modeling wrappers around base R tools are available at https://github.com/PiotrTymoszuk/lmqc.

Clustering of non-CT and non-LF binary clinical features (Appendix 1-table 1) was accomplished with PAM algorithm (partitioning around medoids, package cluster) (Amato et al., 2019) and simple matching distance (SMD, package nomclust) (Boriah et al., 2008) . Association analysis for the participants was performed with a combined procedure involving clustering of the observations by the selforganizing map algorithm (SOM, 4 × 4 hexagonal grid, SMD distance, kohonen package), followed by clustering of the SOM nodes by the Ward.D2 hierarchical clustering algorithm (Euclidean distance, hclust() function, package stats) (Vesanto and Alhoniemi, 2000; Kohonen, 1995; Wehrens and Kruisselbrink, 2018) . Clustering analyses were performed in the participant subset with the complete set of clustering variables. The selection of the optimal clustering algorithm was motivated by the highest ratio of between-cluster to total variance and the best stability measured by mean classification error in 20-fold cross-validation (CV) ( Lange et al., 2004) . The optimal cluster number was determined by the bend of the within-cluster sum-of-squares curve (function fviz_nbclust(), package factoextra) and by the stability in 20-fold CV ( Figure 6 -figure supplement 1C and D, Figure 7 -figure supplement 1D and F; Lange et al., 2004; Wang, 2010) , as well as by a visual inspection of the SOM node clustering dendrograms ( Figure 7-figure supplement 1E ). Assignment of 180-day follow-up outcome features to the clusters of clinical parameters was accomplished with a k-nearest neighbor (kNN) label propagation algorithm (Appendix 1-table 3; Sahanic et al., 2021; Leng et al., 2013) . Cluster assignment visualization in a four-dimensional principal analysis score plot was done with the PCAproj() tool (package pcaPP) (Croux et al., 2007) . To determine the importance of particular clustering variables, the variance (between-cluster to total variance ratio) between the initial cluster structure and the structure with random resampling of the variable was compared, as initially proposed for the random forests ML classifier (Breiman, 2001) . Frequencies of the outcome events in the participant clusters were compared with χ 2 test. In-house-developed association analysis wrappers are available at https:// github.com/PiotrTymoszuk/clustering-tools-2.

ML classifiers C5.0 (package C50) (Quinlan, 1993) , random forests (randomForest) (Breiman, 2001) , support vector machines with radial kernel (kernlab) (Weston and Watkins, 1998) , neural networks (nnet) (Ripley, 2014) , and elastic net (glmnet) (Friedman et al., 2010) were trained to predict the 180day follow-up outcomes employing non-CT and non-LF binary explanatory features (Appendix 1- Kuhn, 2008) . Classifier ensembles were constructed with the elastic net procedure (caretStack() function, caretEnsemble package, Appendix 1-table 4; Deane-Mayer and Knowles, 2019) . Classifier performance in the training cohort and CV was assessed by receiveroperating characteristics (ROCs), Cohen's κ and accuracy (packages caret and vcd, Appendix 1table 5; Fleiss et al., 1969; Kuhn, 2008) . Variable importance measures were extracted from the C5.0 (percent variable usage, c5imp() function, package C50) (Quinlan, 1993) , random forests (Δ Gini index, importance(), package randomForest) (Breiman, 2001) , and elastic net classifiers (regression coefficient β, coef(), package glmnet) (Friedman et al., 2010) .

Participant clustering and ML classifiers trained in the CovILD cohort were implemented in an opensource online pulmonary assessment R shiny app (https://im2-ibk.shinyapps.io/CovILD/; code: https:// github.com/PiotrTymoszuk/COVILD-recovery-assessment-app). Prediction of the cluster assignment based on the user-provided patient data is done by the kNN label propagation algorithm Leng et al., 2013) .

The CovILD study participants (n = 145) were predominantly male (57.8%), age ranging between 19 and 87 years. 77.2% of participants displayed preexisting comorbidity, predominantly cardiovascular and metabolic disease. The cohort included mild (outpatient care, 24.8%), moderate (hospitalization without oxygen supply, 25.5%), severe (hospitalization with oxygen supply, 27.6%), and critical (intensive care unit [ICU] treatment, 22.1%) cases of acute COVID-19 ( Table 1) . The majority of hospitalized participants received anti-infectives during acute COVID-19, anticoagulative, and/or antiplatelet treatment introduced primarily in the ventilated patients. Systemic steroid administration was initiated at the discretion of the physician beginning from week 4 after diagnosis ( Table 2) .

Most patients, irrespective of the acute COVID-19 severity, showed a significant resolution of disease symptoms over time ( Figure 1, Figure 2A ). Persistent complaints at the 6-month follow-up were reported by 49% of the study subjects (Table 3) , with self-reported impaired physical performance (34.7%), sleep disorders (27.1%), and exertional dyspnea (22.8%) as leading manifestations. The frequency of all investigated symptoms declined significantly, even though the pace of their resolution was remarkably slower in the late (100-and 180-day follow-ups) than in the early recovery phase (acute COVID-19 till 60-day follow-up) ( Figure 2B ). Impaired LF was observed in 33.6% of the participants at the 6-month follow-up (Table 3) . Except for the critical COVID-19 survivors (60 days: 66.7%; 180 days post-COVID-19: 50%), no significant reduction in the frequency of LF impairment over time was observed (Figure 3 ). At the 6-month follow-up, structural lung abnormalities were found in 48.5% of patients and moderate-to-severe radiological lung alterations (CT severity score > 5) were present in 19.4% of participants ( Table 3 ). The majority of the participants with impaired LF displayed radiological lung findings. However, a substantial fraction of CT abnormalities, especially mild ones, were accompanied neither by persistent symptoms nor by LF deficits ( The frequency, scoring, and recovery of CT lung findings were related to the severity of acute infection. Pulmonary lesions scored > 5 CT severity points at the 180-day follow-up were most frequent in the individuals with severe and critical acute COVID-19 ( Figure 3-figure supplement 3) . Notably, the hospitalized group with oxygen therapy demonstrated the fastest recovery kinetics. As for the symptom resolution, LF and CT lung recovery decelerated in the late phase of COVID-19 convalescence ( Figure 3 ).

To identify risk factors of delayed recovery at the 6-month follow-up, we screened a set of 52 binary clinical parameters (Appendix 1- -table 2) . By this means, no significant correlates for long-term symptom persistence were identified. Risk factors and readouts of severe and critical COVID-19 including multimorbidity, malignancy, male sex, prolonged hospitalization, ICU stay, and immunosuppressive therapy were significantly associated with persistent CT (Figure 4 ) and LF abnormalities ( Figure 5 ). Persistently elevated inflammatory markers, IL-6 (>7 ng/L) and CRP (>0.5 mg/L), were strong unfavorable risk factors for incomplete radiological and functional pulmonary recovery. Additionally, the biochemical readout of microvascular inflammation, D-dimer (>500 pg/mL) was significantly linked to LF deficits. Low serum anti-S1/S2 IgG titers at the 60-day follow-up and ambulatory acute COVID-19 correlated with an improved pulmonary recovery (Figures 4 and 5) .

Employing the unsupervised PAM algorithm (Amato et al., 2019) , three clusters of co-occurring non-CT and non-LF clinical features of acute COVID-19 and early convalescence (Appendix 1- The 6-month follow-up outcome variables were incorporated in the cluster structure using kNN prediction (Leng et al., 2013) . Long-term symptom persistence was associated with acute and longlasting COVID-19 symptoms in cluster 3, whereas pulmonary outcome parameters were grouped with cluster 2 features ( Figure 6A, Figure 6 -figure supplement 2, Appendix 1- . Kinetic of pulmonary recovery. Recovery from any lung computed tomography (CT) abnormalities, moderate-to-severe lung CT abnormalities (severity score > 5), and recovery from functional lung impairment were investigated in the participants stratified by acute COVID-19 severity by mixedeffect logistic modeling (random effect: individual; fixed effect: time). Significance was determined by the likelihood ratio test corrected for multiple testing with the Benjamini-Hochberg method. Frequencies of the given abnormality at the indicated time points are presented, and p-values and the numbers of complete observations are indicated in the plots.

The online version of this article includes the following figure supplement(s) for figure 3: Figure 4 . Risk factors of persistent radiological lung abnormalities. Association of 52 binary explanatory variables (Appendix 1-table 1) with the presence of any lung computed tomography (CT) abnormalities (A) or moderate-to-severe lung CT abnormalities (severity score > 5) (B) at the 180day follow-up visit was investigated with a series of univariate logistic models (Appendix 1-table 2). Odds ratio (OR) significance was determined by Wald Z test and corrected for multiple testing with the Benjamini-Hochberg method. ORs with 95% confidence intervals for significant favorable and were found the closest cluster neighbors of mild CT abnormalities (severity score ≤ 5). Moderate-tosevere structural alterations (severity score > 5) and LF deficits were, in turn, tightly linked to markers of protracted systemic inflammation (IL-6, CRP, anemia of inflammation) (Sonnweber et al., 2020 ; Figure 6B ).

Risk stratification for perturbed pulmonary recovery by unsupervised clustering Next, we tested whether subsets of patients at risk of an incomplete 6-month recovery may be defined by a similar clustering procedure employing exclusively non-CT and non-LF clinical variables (Appendix 1-table 1). Applying a combined SOM -hierarchical clustering approach, three clusters of the study participants were identified (Figure 7, Figure 7 -figure supplement 1; Vesanto and Alhoniemi, 2000; Kohonen, 1995) . Prolonged hospitalization, anti-infective therapy, overweight or obesity, pain during acute COVID-19, and low anti-S1/S2 titers at the 60-day follow-up were found the most influential clustering features (Figure 7-figure supplement 2; Breiman, 2001) . The patient subsets identified by the SOM approach differed significantly in frequency of radiological lung abnormalities and substantially, yet not significantly, in the frequency of LF impairment at the 180-day unfavorable factors are presented in forest plots. Model baseline (ref) and numbers of complete observations are presented in the plot axis text. Q1, Q2, Q3, Q4: first, second, third, and fourth quartile of anti-S1/S2 IgG titer; ICU: intensive care unit. Odds ratio (OR) significance was determined by Wald Z test and corrected for multiple testing with the Benjamini-Hochberg method. ORs with 95% confidence intervals for the significant favorable and unfavorable factors are presented in a forest plot. Model baseline (ref) and n numbers of complete observations are presented in the plot axis text. Q1, Q2, Q3, Q4: first, second, third, and fourth quartile of anti-S1/S2 IgG titer; CKD: chronic kidney disease. Figure 6 . Association of incomplete symptom, lung function, and radiological lung recovery with demographic and clinical parameters of acute COVID-19 and early recovery. Clustering of 52 non-computed tomography (non-CT) and non-lung function binary explanatory variables recorded for acute COVID-19 or at the early 60-day follow-up visit (Appendix 1-table 1) was investigated by partitioning around medoids (PAM) algorithm with simple matching distance (SMD) dissimilarity measure ( Figure 6-figure supplement 1, Appendix 1-table 3) . The cluster assignment for the Figure 6 continued on next page follow-up. In particular, most of the individuals assigned to the largest, low-risk (LR) subset were CT and LF abnormality-free. The frequency and severity of radiological pulmonary findings were elevated in the smallest intermediate-risk subset (IR) and peaked in the high-risk (HR) group ( Figure 8A) . Despite a comparable frequency of long-term symptoms between the LR, IR, and HR subsets ( Figure 8A ), the HR collective showed the lowest prevalence of dyspnea, cough, night sweating, pain, gastrointestinal manifestations, and complete absence of hyposmia at the 180-day follow-up ( Figure 8B ). Although the LR subset primarily comprised mild COVID-19 cases and the HR subset ICU survivors, the cluster assignment (IR vs. LR, HR vs. LR) remained an independent correlate of persistent CT and LF abnormalities after adjustment for the acute COVID-19 severity ( 

Finally, we investigated if the 6-month follow-up outcome may be predicted by ML classifiers trained with a set of non-CT and non-LF variables recorded during acute COVID-19 and at the 60-day follow-up (Appendix 1-table 1). To this end, five technically unrelated ML classifiers were tested (Appendix 1-table 4; Kuhn, 2008) : C5.0 (Quinlan, 1993) , random forests (RF) (Breiman, 2001) , support vector machines with radial kernel (SVM-R) (Weston and Watkins, 1998) , shallow neural network (Nnet) (Ripley, 2014) , and elastic net generalized linear regression (glmNet) (Friedman et al., 2010) . In addition, the single classifiers with varying outcome-specific accuracy ( Figure 9 figure supplement 1) were bundled into ensembles by the elastic net procedure (Figure 9 -figure supplement 2, Appendix 1-table 4; Kuhn, 2008; Deane-Mayer and Knowles, 2019) . Finally, the classifier and ensemble performance was investigated in the training cohort and 20-fold CV by ROC (Appendix 1-table 5).

All tested ML algorithms and ensembles demonstrated good accuracy (area under the curve [AUC] > 0.78) and sensitivity (>0.84) at predicting any lung CT abnormalities at the 6-month follow-up in the study cohort serving as a training data set. Their efficiency in CV was moderate (AUC: 0.69-0.81; sensitivity: 0.69-0.78) (Figure 9 , Figure 9 -figure supplement 3, Appendix 1-table 5). In turn, moderate-to-severe structural lung findings were recognized with markedly lower sensitivity both in the training data set (>0.43) and the CV (0.39-0.48). Even though impaired LF and persistent symptoms were common at the 6-month follow-up in the training data set (Figures 2 and 3) , nearly half of the cases were not identified by any of the tested ML algorithms and their ensembles in the CV setting ( Figure 9 , Figure 9 -figure supplement 3, Appendix 1-table 5). The sensitivity of the ensembles and single classifiers at predicting CT and LF abnormalities was substantially better in severe and critical COVID-19 survivors than in ambulatory and moderate cases ( Figure 10 , Appendix 1-table 6).

The most important explanatory variables for pulmonary abnormalities by three unrelated classifiers (C5.0, RF, and glmNet) included preexisting malignancy, multimorbidity, markers of systemic inflammation (IL-6 and CRP), and anti-S1/S2 antibody levels at the 60-day follow-up ( outcome variables at the 180-day follow-up visit (persistent symptoms, functional lung impairment, mild lung CT abnormalities [severity score ≤ 5] and moderate-to-severe lung CT abnormalities [severity score > 5]) was predicted by k-nearest neighbor (kNN) label propagation procedure. Numbers of complete observations and numbers of features in the clusters are indicated in (A). (A) Cluster assignment of the outcome variables (diamonds) presented in the plot of principal component (PC) scores. The first two major PCs are displayed. The explanatory variables are visualized as points. Percentages of the data set variance associated with the PC are presented in the plot axes. (B) Five nearest neighbors (lowest SMD) of the outcome variables presented in radial plots. Font size, point radius, and color code for SMD values. Q1, Q2, Q3, Q4: first, second, third, and fourth quartile of anti-S1/S2 IgG titer; GITD: gastrointestinal disease; CKD: chronic kidney disease; ICU: intensive care unit; COPD: chronic obstructive pulmonary disease.

The online version of this article includes the following figure supplement(s) for figure 6: Figure 7 . Clustering of the study participants by non-lung function and non-computed tomography (non-CT) clinical features. Study participants (n = 133 with the complete variable set) were clustered with respect to 52 non-CT and non-lung function binary explanatory variables recorded for acute COVID-19 or at the 60-day follow-up visit (Appendix 1-table 1) using a combined self-organizing map (SOM: simple matching distance) and hierarchical clustering (Ward.D2 method, Euclidean distance) procedure ( 

Herein, we prospectively evaluated trajectories of COVID-19 recovery in an observational cohort enrolled in the Austrian CovILD study . Despite the resolution of symptoms and pulmonary abnormalities at the 6-month follow-up in a large fraction of the study participants, the recovery pace was substantially slower in the late convalescence when compared with the first three months after diagnosis Huang et al., 2021a) . Persistent symptoms and CT findings were detected in more than 40% and reduced LF in approximately one-third of the cohort, which is in line with recovery kinetics and signs of lung lesion chronicity reported by others (Caruso et al., 2021; Huang et al., 2021b; Huang et al., 2021a; Faverio et al., 2021; Hellemons et al., 2021; Zhou et al., 2021) . By comparison, similar protracted pulmonary recovery was reported for SARS (Hui et al., 2005; Ng et al., 2004; Ngai et al., 2010; Lam et al., 2009 ) and non-COVID-19 acute respiratory distress syndrome (Wilcox et al., 2013; Masclans et al., 2011) . Of note, treatment approaches for hospitalized patients in our cohorts and similar cohorts recruited at the pandemic onset in early 2020 (Caruso et al., 2021; Huang et al., 2021b; Huang et al., 2021a; Faverio et al., 2021; Hellemons et al., 2021) differ significantly from the current standard of care for acute COVID-19, which includes early systemic steroid use and antiviral and various immunomodulatory medications. How improved standardized therapy and anti-SARS-CoV-2 vaccination affect the clinical and pulmonary recovery needs to be investigated.

In roughly half of our study participants with abnormal lung CT findings, and especially in those with low-grade structural abnormalities, no overt LF impairment at follow-up was discerned. Still, even subclinical lung alterations may bear the potential for clinically relevant progression of interstitial lung disease (Suliman et al., 2015; Hatabu et al., 2020) requiring systematic CT and LF monitoring. Conversely, symptom persistence was weakly associated with incomplete functional or structural pulmonary recovery.

Since PASC are found in as many as 10% of COVID-19 patients Venkatesan, 2021; Sudre et al., 2021b) , robust, resource-saving tools assessing the individual risk of pulmonary complications are urgently needed (Shah et al., 2021; Raghu and Wilson, 2020) . Covariates and characteristics of severe acute COVID-19 such as male sex, age, and preexisting comorbidities, hospitalization, ventilation, and ICU stay were proposed as the risk factors of persistent pulmonary impairment Caruso et al., 2021; Huang et al., 2021a; Faverio et al., 2021; Raghu and Wilson, 2020) . However, their applicability in predicting complications of pulmonary recovery from mild or moderate COVID-19 is limited. Our results of univariate modeling, clustering, and ML prediction point towards a distinct long-term pulmonary risk phenotype that manifests during acute COVID-19 and early recovery and whose central components are protracted systemic (IL-6, CRP, anemia of inflammation) and microvascular inflammation (D-dimer), and strong humoral response (anti-S1/S2 IgG) demographic risk factors and comorbidities (Sonnweber et al., 2020) . Hence, consecutive monitoring of systemic inflammatory parameters analogous to concepts of interstitial lung disease in autoimmune disorders (Khanna et al., 2020) and anti-S1/S2 antibody levels may improve identification of the individuals at risk of chronic pulmonary damage irrespective of the acute COVID-19 severity.

Clustering and ML have been employed for deep phenotyping and predicting acute and postacute COVID-19 outcomes in multivariable data sets Sudre et al., 2021a; Estiri et al., 2021; Demichev et al., 2021; Benito-León et al., 2021) . We demonstrate that subsets of COVID-19 patients that significantly differ in the risk for long-term CT abnormalities may be defined by an easily accessible clinical parameter set available at the early post-COVID-19 assessment. This approach did not involve any CT or LF variables. Furthermore, the cluster classification correlated with axes. (B) Presence of the most influential clustering features (Figure 7-figure supplement 2) in the participant clusters presented as a heat map. Cluster #1, #2, and #3 refer to the feature clusters defined in Figure 6 . Q1, Q2, Q3, Q4: first, second, third, and fourth quartile of anti-S1/S2 IgG titer; GITD: gastrointestinal disease; CKD: chronic kidney disease; CVD: cardiovascular disease; GI: gastrointestinal; PD: pulmonary disease.

The online version of this article includes the following figure supplement(s) for figure 7: the risk of long-term pulmonary abnormalities independently of the acute COVID-19 severity. Thus, these characteristics provide a useful tool for broad screening of convalescent populations, including individuals who experienced mild or moderate COVID-19.

We show that technically unrelated ML classifiers and their ensemble trained without CT and LF explanatory variables can predict lung CT findings independently of their grading at the 6-month follow-up with good specificity and sensitivity in the training collective and CV. By contrast, the more specific prediction of moderate-to-severe lung CT or risk estimation for LF deficits demonstrated a limited sensitivity. For the moderate-to-severe CT abnormalities, this can be primarily traced back to their low frequency resulting in a suboptimal classifier training, especially in CV. A substantial fraction of the participants (20.7%, n = 30) suffered from a preexisting respiratory condition (pulmonary disease, asthma, or COPD) likely paralleled by LF reduction, which possibly confounded the prediction of the post-COVID-19 LF deficits both by clustering and ML. Accumulating evidence suggests that post-acute COVID-19 symptoms are highly heterogeneous conditions with multiorgan, neurocognitive, and psychological manifestations Evans et al., 2021; Davis et al., 2021) , which may differ in risk factor constellations. This could explain why univariate modeling, clustering, and ML failed to estimate persistent symptom risk in our small study cohort. In general, the ML prediction quality may greatly benefit from a larger training data set and inclusion of additional explanatory variables such as cellular readouts of inflammation, in-depth medication, and broader acute symptom data. Nevertheless, the herein described cluster-and ML classifiers represent resourceeffective tools that may assist in the screening of medical record data and identification of COVID-19 patients requiring systematic CT and LF monitoring. To facilitate the identification of patients at risk for protracted respiratory recovery and enable validation in an external collective, we implemented the clustering and prediction procedures in an open-source risk assessment application (https://im2ibk.shinyapps.io/CovILD/).

Our study bears limitations primarily concerning the low sample size and the cross-sectional character of the trial. Because of the impaired availability of the patients and the prolonged inpatient rehabiliation, the 60-and 100-day follow-up visits in part showed a temporal overlap that may have impacted the accuracy of the longitudinal data. Missingness of the consecutive outcome variable record and the participant dropout, particularly of mild and moderate COVID-19 cases, may have also potentially confounded the participant clustering results and ML risk estimation for CT abnormalities and LF impairment since prolonged hospitalization was found to be a crucial cluster-defining and influential explanatory feature. Additionally, even though the reproducibility of the risk assessment algorithms was partially addressed by CV, cluster and ML classifiers call for verification in a larger, independent multicenter collective of COVID-19 convalescents.

In summary, in our CovILD study cohort we found a high frequency of CT and LF abnormalities and persistent symptoms at the 6-month follow-up, and a flattened recovery kinetics after 3 months post-COVID-19. Systematic risk modeling reveled a set of clinical variables linked to protracted pulmonary recovery apart from the severity of acute infection such as inflammatory markers, anti-S1/S2 IgG levels, multimorbidity, and male sex. We demonstrate that clustering and ML classifiers may help to identify individuals at risk of persistent lung lesions and to relocate medical resources to prevent longterm disability. . Prediction of persistent radiological lung abnormalities, functional lung impairment, and symptoms by machine learning algorithms. Single machine learning classifiers (C5.0; RF: random forests; SVM-R: support vector machines with radial kernel; NNet: neural network; glmNet: elastic net) and their ensemble (Ens) were trained in the cohort data set with 52 non-computed tomography (non-CT) and non-lung function binary explanatory variables recorded for acute COVID-19 or at the 60-day follow-up visit (Appendix 1-table 1) for predicting outcome variables at the 180-day follow-up visit (any lung CT abnormalities, moderate-to-severe lung CT abnormalities [severity score > 5], functional lung impairment, and persistent symptoms) (Appendix 1-table 4). The prediction accuracy was verified by repeated 20-fold cross-validation (five repeats). Receiver-operating characteristics (ROCs) of the algorithms in the cross-validation are presented: area under the curve (AUC), sensitivity (Sens), and specificity (Spec) (Appendix 1-table 5). The numbers of complete observations and outcome events are indicated under the plots.

The online version of this article includes the following figure supplement(s) for figure 9: Figure 10 . Performance of the machine learning ensemble classifier in mild-to-moderate and severe-to-critical COVID-19 convalescents. The machine learning classifier ensemble (Ens) was developed as presented in Figure 9 . Its performance at predicting outcome variables at the 180-day follow-up visit (any computed tomography [CT] lung abnormalities, moderate-to-severe lung CT abnormalities [severity score > 5], functional lung impairment, and persistent symptoms) in the entire cohort, mild-to-moderate (outpatient or hospitalized without oxygen), and severe-to-critical COVID-19 convalescents (oxygen therapy or ICU) in repeated 20-fold cross-validation (five repeats) was assessed by receiver-operating characteristic (ROC) (Appendix 1- 

The complete R analysis pipeline and the anonymized study data in form of stratified study variables are available as a public GitHub repository: https://github.com/PiotrTymoszuk/CovILD_6_Months (copy archived at swh:1:rev:df521ede1d284e074a0484d3e4d0ce71097d00c3). The R code for the key tools used for uni-variate modeling and model quality control (Figures 4 and 5 , https://github.com/ PiotrTymoszuk/lmqc; copy archived at swh:1:rev:a020119d8f23b60901115c5c2ce6f6c71998ed31), cluster analysis and its quality control (Figures 6-7 , https://github.com/PiotrTymoszuk/clusteringtools-2; copy archived at swh:1:rev:64141197ca28838a8978dce9093443537157d79f) and the risk assessment applicaiton (https://github.com/PiotrTymoszuk/COVILD-recovery-assessment-app; copy archived at swh:1:rev:95f02215f4c13425d3b76f6a13b7862a53279ab9) is available at GitHub. Source data for Figures 2-10 has been included as Source data 1. 

Similarity Search and Applications. Faster K-Medoids Clustering: Improving the PAM

Fitting linear mixed-effects models using lme4

Using Unsupervised Machine Learning to Identify Age-and Sex-Independent Severity Subgroups Among Patients with COVID-19: Observational Longitudinal Study

Controlling the False Discovery Rate: A Practical and Powerful Approach to Multiple Testing

Similarity Measures for Categorical Data: A Comparative Evaluation

Random forests

Post-Acute Sequelae of COVID-19 Pneumonia: Six-month Chest CT Follow-up

Algorithms for Projection-Pursuit robust principal component analysis

Characterizing long COVID in an international cohort: 7 months of symptoms and their impact

A time-resolved proteomic and prognostic map of COVID-19

An interactive web-based dashboard to track COVID-19 in real time

Consortium for Characterization of COVID-19 by EHR (4CE). 2021. Evolving phenotypes of non-hospitalized patients that indicate long COVID

Physical, cognitive, and mental health impacts of COVID-19 after hospitalisation (PHOSP-COVID): a UK multicentre, prospective cohort study. The Lancet

Six-Month Pulmonary Impairment after Severe COVID-19: A Prospective, Multicentre Follow-Up Study. Respiration

Harmonization of six quantitative SARS-CoV-2 serological assays using sera of vaccinated subjects

Large sample standard errors of kappa and weighted kappa

Regularization Paths for Generalized Linear Models via Coordinate Descent

Interstitial lung abnormalities detected incidentally on CT: a Position Paper from the Fleischner Society. The Lancet

Ribbers GM, van den Berg-Emons HG

A Longitudinal Study of Respiratory, Physical and Psychological Outcomes

Clinical features of patients infected with 2019 novel coronavirus in Wuhan

6-month consequences of COVID-19 in patients discharged from hospital: a cohort study

1-year outcomes in hospital survivors with COVID-19: a longitudinal cohort study

The 1-year impact of severe acute respiratory syndrome on pulmonary function, exercise capacity, and quality of life in a cohort of survivors

Johns Hopkins Coronavirus Resource Center. 2021. COVID-19 Map

Etiology, Risk Factors, and Biomarkers in Systemic Sclerosis with Interstitial Lung Disease

Self-Organizing Maps

Building predictive models in R using the caret package

lmerTest Package: Tests in Linear Mixed Effects Models

Mental morbidities and chronic fatigue in severe acute respiratory syndrome survivors: long-term follow-up

Stability-based validation of clustering solutions

Adaptive Semi-Supervised Clustering Algorithm with Label Propagation

Quality of life, pulmonary function, and tomographic scan abnormalities after ARDS

Six month radiological and physiological outcomes in severe acute respiratory syndrome (SARS) survivors

The long-term impact of severe acute respiratory syndrome on pulmonary function, exercise capacity and health status

Serology-informed estimates of SARS-CoV-2 infection fatality risk in Geneva

nVenn: generalized, quasi-proportional Venn and Euler diagrams

C4.5: Programs for Machine Learning

COVID-19 interstitial pneumonia: monitoring the clinical course in survivors. The Lancet

Pattern Recognition and Neural Networks

plotROC: A Tool for Plotting ROC Curves

Phenotyping of acute and persistent COVID-19 features in the outpatient setting: exploratory analysis of an international cross-sectional online survey

Managing the long term effects of covid-19: summary of NICE, SIGN, and RCGP rapid guideline

Persisting alterations of iron homeostasis in COVID-19 are associated with nonresolving lung pathologies and poor patients' performance: a prospective observational cohort study

Cardiopulmonary recovery after COVID-19: an observational prospective multicentre trial

Symptom clusters in COVID-19: A potential clinical prediction tool from the COVID Symptom Study app

Attributes and predictors of long COVID

Brief Report: Pulmonary Function Tests: High Rate of False-Negative Results in the Early Detection and Screening of Scleroderma-Related Interstitial Lung Disease. Arthritis & Rheumatology

NICE guideline on long COVID. The Lancet

Clustering of the self-organizing map

Consistent selection of the number of clusters via crossvalidation

Flexible self-organizing maps in kohonen 3.0

Multi-Class Support Vector Machines

Ggplot2: Elegant Graphics for Data Analysis

Welcome to the Tidyverse

Radiologic outcomes at 5 years after severe ARDS

Fundamentals of Data Visualization: A Primer on Making Informative and Compelling Figures

We acknowledge the commitment of the staff and providers of our institutions through the COVID-19 crisis and the suffering and loss of our patients as well as their families. PT is (from May 2021 on) a freelance data scientist working in his own enterprise 'Data Analytics as a Service Tirol'. He received an

Appendix 1-