Predicting the persistence of Long COVID 36 months after hospitalisation: a prospective Ukrainian cohort study

Objective To determine early predictive markers and to develop prognostic models for self-reported persistence of Long COVID syndrome at 36 months after hospitalisation. Design A prospective observational study with cross-sectional analyses at predefined follow-up points. Setting A tertiary care hospital. Participants 221 patients hospitalised for COVID-19 underwent detailed clinical, functional, imaging and survey-based symptom assessment prior to discharge and at 1 month, followed by a 3-year longitudinal follow-up on symptoms. 159 patients (mean age 53.1±13.6 years, 55% female) completed the 36-month assessment and were included in the final analysis. Main outcome measure Self-reported persistence of Long COVID symptoms at 36 months after hospital discharge. Results At 36 months after discharge, 66 (41%) patients reported persistent Long COVID symptoms. This group did not differ in age or sex structure, more frequently reported ongoing symptoms at 3 months (91% vs 59%, p<0.001) and 12 months (77% vs 13%, p<0.001), had longer time to hospitalisation (10.2±3.4 vs 5.2±3.9 days, p<0.001) and higher Charlson comorbidity index (0.61±0.96 vs 0.33±0.56, p=0.022), higher in-hospital creatinine (109 (100–119) vs 96 (85–105) µmol/L, p<0.001), aspartate aminotransferase (AST) (43 (37–55) vs 32 (21–47) IU/L, p=0.003), and alanine aminotransferase (47 (30–128) vs 42 (26–62) IU/L, p=0.011) and paradoxically lower Galectin-3 (8.6±0.7 vs 9.0±0.5 ng/mL, p<0.001) and sST2 (23.2±13.7 vs 29.7±15.8 ng/mL, p=0.012). No significant differences were found in inflammatory markers, minimal in-hospital peripheral capillary oxygen saturation, need for oxygen supplementation, pulmonary involvement on CT or lung ultrasound, and clinically negligible differences detected in cardiovascular evaluation, with a baseline tendency to minimal hypercontractility and lower flow-mediated dilation (10.1±5.1 vs 12.9%±5.9%, p=0.006) and marginally worse early dynamics of diastolic indices in Long COVID group. In 6-Minute Walking Test, no differences were observed at baseline but patients with Long COVID reported higher postexertional fatigue (3.3±2.1 vs 1.9±1.6, p<0.001) and dyspnoea (3.4±1.7 vs 2.5±1.8, p=0.027) despite lower blood oxygen desaturation (1.7±1.7 vs 2.8%±2.4%, p=0.014) at 1 month. In symptom assessment, longitudinal trajectories differed significantly for Medical Research Council Dyspnoea (p=0.004), post-COVID functional scale (p<0.001), EFTER-COVID study physical symptoms (p<0.001) and Symptom Burden Questionnaire for Long COVID Memory, Thinking and Communication (p<0.001) scores, reaching maximal differences by 3–12 months, while respiratory scores showed no sustained differences. A simple artificial neural network model combining predischarge variables (sex, age, creatinine, AST, weight loss, duration of hospitalisation and 6-Minute Walk Distance) with 3-month multidimensional symptom assessment achieved a 92% accuracy (area under the receiver operating characteristic, AUROC 0.902), consistently outperforming other machine learning classifiers. A simplified model using only 3-month survey data, age, sex and hospitalisation duration retained high discrimination accuracy (88%, AUROC 0.850). Conclusions 41% of hospitalised COVID-19 patients reported persistent symptoms at 36 months, preceded by early divergence in patient-reported symptoms despite comparable recovery of objective functional parameters. Predictive models incorporating 3-month multidimensional symptom assessment and routine clinical variables accurately identified patients at risk, outperforming later reassessment strategies and supporting early risk stratification.