A total of 44 subjects (31 men and 13 women) were initially enrolled in this non-randomized experimental prospective study. Two groups were formed throughout study: the treatment group (30 subjects, 56.03 ± 14.05 years old) and the dropout group (12 subjects, 56.33 ± 11.33 years old), which consisted of subjects who dropped out of the program (see Figure 1).

To be eligible for the study, participants had to meet the following criteria: moderate presentation of COVID-19 illness (patients hospitalized in intermediate care units without ICU criteria) or severe presentation of COVID-19 illness (patients requiring ventilatory support in intensive care); elapsed time from hospital discharge less than 1.5 years; age above 18 years; adequate level of comprehension (Mini Mental State Examination, MMSE >26 points); chronic pathologies under pharmacological treatment and control; literacy; absence of suicidal ideation (according to C-SSRS Baseline-Screening); seeking rehabilitation services for the first time; and signed informed consent in writing for participation in the study. Participants with pre-existing cardiovascular, metabolic, and respiratory diseases were evaluated by the medical teams to ensure secure control and stabilization.

The study obtained approval from the Ethics Committee in Scientific Research of the University of Antofagasta (CEIC-UA) (approval #339/2021) and complies with the standards specified in the Nuremberg Code, Declaration of Helsinki, CIOMS, and guidelines of Ezekiel Emanuel. The study was conducted at the Physical Medicine and Rehabilitation Unit of the Clinical Hospital of the University of Antofagasta, Chile, and involved personnel from the University of Antofagasta. The University of Antofagasta Clinical Hospital, established in 2019 through a collaboration between the University of Antofagasta and the Antofagasta Regional Government, serves the dual purpose of advancing specialist training, addressing the scarcity of such professionals in the region, and creating a clinical environment for educating medical healthcare sciences students across diverse specialties. Furthermore, it offers training opportunities for professionals in fields like Medical Technology, Nursing, Obstetrics, Kinesiology, and Occupational Therapy.

A call was made to the Antofagasta community through the social networks of the Clinical Hospital of the University of Antofagasta, as well as through television and radio in the city.

Participants who met the eligibility criteria entered the study after a medical anamnesis and pharmacological assessment to determine the appropriate rehabilitation for their needs (see Figure 2).

The participants underwent an 8-week treatment program with 100% attendance compliance. The treatment included: physical therapy twice a week, speech and language therapy twice a week, psychological therapy once a week, and nutritional consultation once a month (see Figure 3).

The order of administration assessments was randomized, and participants were instructed to request rest breaks as needed. Trained investigators from the Physical Medicine and Rehabilitation Service of HCUA, together with last-year university students in practice at the University of Antofagasta, conducted data collection.

Vital signs of the patients were monitored by medical staff and students during clinical assessments and therapies.

We followed the TIDieR (Template for Intervention Description and Replication) checklist to provide a comprehensive and standardized description of our rehabilitation intervention (20).

The quality of life was assessed using the Spanish version of the Short Form-36 Health Survey (SF-36) questionnaire (21), which is used in the Public Health System in Chile (22). The questionnaire consists of 36 items that cover various scales related to: physical function, physical role, body pain, general health, vitality, social function, emotional role, and mental health. The questionnaire was self-administered by the patients in a quiet room of the hospital following the methodology described by Vilagut et al. (21). To consider a scale as valid, patients were required to have answered at least 50% of its items. To calculate the score, any unanswered item was substituted with the average value derived from the completed items.

Functionality was estimated using the post-COVID-19 Functional Status (PCFS) Scale (23), a Spanish version recommended by Scientific Societies and Professional Colleges of the Rehabilitation Area (15). The scale was applied through a structured interview conducted by trained researchers at the beginning and end of the 8-weeks rehabilitation program.

The sample size was calculated using the software G*Power 3.1, considering the Wilcoxon test for two related samples to analyze the changes between baseline and 8 weeks of intervention. A bilateral contrast, an effect size of d = 0.5, a significance level of α = 0.05, and a power of 0.80 resulted in a minimum sample size of 34. To account for potential variables with smaller effect sizes, we increased the sample size to 44.

All statistical analyses were performed using MATLAB (MATLAB_R2022a, 2022). The significance level was set at α = 0.05 (two-sided) for all tests. Data are reported as means and standard errors [mean (SD)] or as numbers and percentages [n (%)].

The Shapiro-Wilk test was used to assess the data distribution, and most of the data was found to follow a normal distribution. For normally distributed data, the differences between baseline and 8 weeks of interventions were evaluated using the paired Student's t-test. Non-normally distributed variables were analyzed, using the non-parametric Wilconxon sign-rank test for paired samples and Pearson chi-square (X²) for proportions. In cases where two variables had different distributions, the non-parametric test was chosen. The effect size of the differences between groups and conditions was computed as Cohen's-d by Ruggero G. Bettinardi (46).

Logistic regression models were employed to predict the responder type: good or poor. A good responder was defined as an individual who experienced an increase equal to or greater than 30% in the delta of the SF-36 total score, which was calculated using the formula [(end—start)/end] * 100. The predictors considered in the analysis were age, sex, COVID-19 category, length of hospital stay, time elapsed after hospital discharge, pre-hospitalization vaccinations, post-hospitalization vaccinations, arterial hypertension (HT), diabetes mellitus (DM), dyslipidemia, depression, respiratory disease (asthma or chronic obstructive pulmonary disease), smoking status, and alcohol consumption.

The delta of SF-36 total score for patients who underwent rehabilitation was predicted using a multivariate linear regression model with backward selection. This approach initially included all variables in the model and subsequently removed the least significant ones (p ≥ 0.05) until all remaining variables were significant in explaining the outcome. The predictors considered in the analysis were the deltas of changes in the nine areas assessed by the SF-36 questionnaire, which encompassed physical function, physical role, body pain, general health, vitality, social role, emotional role, mental health, and health transition.

Finally, the Classification Learner App available in the Matlab Statistics and Machine Learning Toolbox (Mathworks, Natick MA), was employed to develop a classification model able to predict dropouts. A comprehensive set of 86 predictors was considered, including general data variables (such as age, sex, days of hospitalization, among others) and baseline variables (pre-intervention outcomes). The Chi2 algorithm for feature ranking (47) was employed to select the top 10 highest-ranking features.

Several classifiers, including Coarse tree, Medium Tree, and Fine Tree, were evaluated, with Coarse tree yielding the best results. The Coarse tree classifier was trained using different values for the cost matrix, enabling the assignment of different misclassification costs to each type of error. This approach is particularly crucial in imbalanced datasets, such as the one in this study, where the number of subjects in the treatment group is almost three time higher than the number of subjects in the dropout groups (positive group) Therefore, false negatives (FN) were penalized with a cost of 3, and false positives (FP) were penalized with a cost of 1. This penalty scheme aimed to optimize the model's performance in classifying the minority class. Furthermore, 5-fold cross-validation was performed to assess the model's performance. The model underwent 100 iterations, each trained using the fitctree function and the top 10 highest-ranking features determined by fscchi2. The accuracy of these 100 iterations was calculated, and the mean value, along with the standard deviation (SD) was reported. It is worth noting that the fitctree function excludes observations with entirely missing values for the predictor vector. However, observations with some missing values are used to find splits on variables where those observations have valid values (48).

Our study examined participants from northern Chile who had experienced severe COVID-19 revealing distinct morbidity characteristics. Notably, there was a prevalence of 55.17% for hypertension and a 20.68% for diabetes mellitus among participants. Additionally, 31.03% of the participants had a history of respiratory disease. Their average length of stay in the intensive care unit (ICU) was 54.89 days, and their average body mass index (BMI) was 32.40 kg/m².

These characteristics differ from other studies. For instance, a study conducted by Chadli (42) at Ibn Rochd CHU (Casablanca, Morocco) reported a lower proportion of hypertension (34.1%) and respiratory disease (4.9%), but a higher prevalence of diabetes mellitus (43.9%). Moreover, they observed a shorter average ICU stay (8.42 days) and a lower average BMI (25.34 kg/m²). Another study conducted in Leuven, Belgium (8), implemented a multidisciplinary respiratory rehabilitation program for 22 patients after COVID-19 hospitalization, which included group sessions over 12 weeks. In this study, only 23% of the patients had a history of pulmonary disease, and the average hospital stay was 29 days, both lower than in our study.

In contrast, our study aligns with Raman et al. (3), where individuals with severe COVID-19 cases displayed a similar average BMI of 32.40 kg/m². Raman et al. (3) examined fifty-eight survivors of moderate to severe COVID-19 infection who were discharged from Oxford University Hospitals National Health Service Foundation Trust (Banbury, Oxfordshire) after 2–3 months. However, their study showed a lower prevalence of hypertension (37.9%) and diabetes mellitus (15.5%), but a higher proportion of patients with a history of respiratory disease (39.7%) compared to our results. Additionally, they had a shorter median ICU stay, lasting only 8.5 days. Notably, a common point among these studies was the higher prevalence of men among severe COVID-19 cases, which aligns with our data showing 68.96% male participants (3, 5, 8, 42).

A population more similar to our study could be expected from the research conducted by Imamura et al. (16) in São Paulo, Brazil. However, in their retrospective case series (n = 27) of patients who received intensive inpatient rehabilitation, they found an average BMI of 27.57 kg/m², which is considerably lower than what we observed in our study. Additionally, the average length of stay in the ICU was shorter (30.04 days).

The unique morbidity characteristics of severe COVID-19 patients in northern Chile underscore the importance of regional variability in impact of the disease. Further research is warranted to comprehend the underlying factors influencing these differences in patient profiles.

Among COVID-19 survivors, it has been documented that one of the most common post-ICU referrals is to a psychologist or psychiatrist (22%) (7). The psychological and emotional context of post-COVID-19 survivors can vary depending on the population under study. For example, Chadli (42) observed that 14.6% of participants experienced moderate to severe depression, while 12.2% reported severe anxiety. In the study of Pistarini et al. (40) on cognitive and emotional disturbances caused by COVID-19 (n = 20) in the rehabilitation unit of San Raffaele Hospital in Milan, Italy, approximately 40% of patients exhibited symptoms of mild to moderate depression. In the Latin American study conducted by Imamura et al. (16), the percentages of moderate to severe depression and anxiety were 0% and 13.64%, respectively. In contrast, our study found higher levels of moderate to severe depression (84.61%) and moderate to severe anxiety (84.61%). These differences highlight the importance of local demographic and clinical studies, as the context of each population and country can influence individual characteristics and impact the results of a multidisciplinary rehabilitation program.

In this context, it is possible that the demographic and clinical characteristics of our population may explain some of our results. Non-continuity in post-COVID rehabilitation programs has been observed before (16), where the main reason was voluntary discharge. In our study, twelve patients discontinued the program before 8 weeks for various reasons, including repeated absences, lack of motivation, loss of telephone follow-up, among others. Explaining this phenomenon, which has multiple facets and occurs in the context of a pandemic, is complex. One possible explanation could be associated with our population of obese survivors, who were likely obese before their COVID-19 infection. Obesity was one of the most common comorbidities observed in severe cases of SARS-CoV-2 (8, 50), and it has been linked to poor attendance rates and compliance, hindering treatment effectiveness. The presence of barriers to behavior change, such as poor motivation, lack of time, health and physical limitations, negative thoughts/moods, socioeconomic factors, lack of enjoyment of exercise, and other determinants of adherence, is a characteristic of this disease. On the other hand, the high prevalence of moderate to severe depression and anxiety among our participants is often accompanied by symptoms commonly observed in the post-illness stage, such as frequent recall of traumatic memories, insomnia, and emotional lability (40, 51). These symptoms have been associated with adherence issues and persistence in therapies (52).

One possible solution to address the issue of patients dropping out during rehabilitation programs is to implement an initial screening to identify individuals with a higher likelihood of discontinuing their participation. Such an approach could help saving both human and structural resources while also enabling the offer of more suitable treatment alternatives tailored to individual preferences and contexts. In line with this strategy, our study introduces a preliminary classification model designed to predict individuals at risk of dropping out. This model incorporates the most influential predictors for treatment adherence, all of which were assessed at baseline.

These predictive variables encompass factors such as the level of dyspnea at the end of the 6MWT, Tinetti balance score, percentage achievement in Tinetti balance, levels of fatigue and dyspnea at the end of the 1'STS, vitality and physical role scores from the SF36 questionnaire, as well as the presence of comorbidities such as DM, HTA, and dyslipidemia. It is noteworthy that these characteristics align with observations from various studies involving post-COVID patients across different geographical locations, where the presence of comorbidities like HT and, DM has also been noted (3, 8, 17, 42).

Additionally, our findings suggest a linear relationship between the duration of rehabilitation and the physical (muscular strength, exercise capacity) and functional improvements (gait, balance, daily activities, among others) achieved by the end of the rehabilitation program (16). On the other hand, the research by Imamura et al. (16) indicated that baseline psychological and cognitive functions at admission did not significantly influence the duration of rehabilitation interventions or the functional outcomes achieved upon discharge. Consequently, these variables were not considered as predictors in our dropout model.

An important aspect of our proposed model is its reliance on low-cost, validated tools that can be readily employed in most healthcare settings. These tools include the 6MWT, 1'STS, SF36 questionnaire, and a comorbidity interview, making the model both practical and accessible for widespread implementation.

Furthermore, our regression model aimed at predicting changes in the SF-36 total score was influenced by four key components of quality of life: physical function, physical role, general health, and mental health. In a study conducted by Líška et al. (53) where the SF-36 questionnaire was administered to long-COVID patients (n = 469), the following components showed mean scores of 66.2 ± 25.4 in physical function, 34.1 ± 21.4 in physical role, 35.8 ± 16.1 in general health, and 38.6 ± 16.0 in mental health. When compared to a control group, all these components showed significant differences. In our patient sample, at baseline conditions, we observed a lower mean score for physical function (59.6 ± 19.7), while higher mean scores were reported for physical role, general health, and mental health (39.0 ± 23.4, 58.8 ± 19.7, 58.8 ± 118.2, respectively). This suggests an opportunity to focus therapeutic efforts on these specific components, potentially reducing costs associated with therapies targeting other aspects of quality of life of COVID-19 survivors.

In our study, demographic and clinical history characteristics were not predictive of whether individuals would be good or poor responders to the multidisciplinary rehabilitation program. This may be due to therapy response being influenced by factors such as sociocultural and socioeconomic variables, which were not considered into account in this study but should be considered in the rehabilitation of COVID-19 survivors (5). It has been observed that the presence of risk factors, such as living in overcrowded conditions, precarious housing, lack of accessible medical care, and employment in higher-risk environments (54), can influence the therapy response of more vulnerable patients, including those within the public healthcare system. While a more comprehensive sociodemographic study of the population in northern Chile who survived COVID-19 is needed, these preliminary findings provide initial insights into population behavior and factors to consider in a multidisciplinary rehabilitation program.

Despite the notable improvements in functional outcomes that underscore the potential of a comprehensive multidisciplinary rehabilitation program, this prospective study does have limitations. These, include a relatively small sample size and the absence of a control group. Additionally, we acknowledge the differences in group sizes, as therapy was customized to meet the individual needs of COVID-19 survivors based on recommendations. However, it is worth to note that this study aims to document the experiences of a rehabilitation service during a pandemic period when there was limited scientific evidence from randomized controlled trials. This is particularly significant in the context of patients reliant on the public healthcare system in Chile or with limited resources, highlighting the relevance of this manuscript for developing countries.