Phenotype evolution in post-COVID-19 chronic cough and prediction of pulmonary fibrosis risk: a longitudinal cohort study leveraging multi-omics integration and deep learning

Post-coronavirus disease 2019 (COVID-19) chronic cough (PCC) is a frequent sequela, affecting approximately 20–30% of patients beyond 8 weeks, with a subset progressing to pulmonary fibrosis (PF). Most existing studies rely on static indicators, which limits monitoring of dynamic disease progression. Therefore, this study aimed to construct longitudinal trajectories of PCC clinical phenotypes and develop an early warning model for PF risk through the integration of multimodal data and deep learning. In a retrospective cohort of 130 patients with PCC (100: training set and 30: validation set), we integrated cough diaries, fractional exhaled nitric oxide (FeNO), chest computed tomography (CT) radiomic features, and serum IL-6 and KL-6 levels. Patients were grouped into three phenotypes using Uniform Manifold Approximation and Projection and consensus clustering: eosinophilic-inflammatory type (FeNO ≥ 40 ppb; IL-6 ≥ 12 pg/mL), neuro-hypersensitive type [nocturnal/diurnal cough-frequency ratio ≥ 1.8; Cough Evaluation Test (CET) average item score ≥ 4], and fibrosis-driven type (annual fibrosis-volume growth ≥ 5% on serial CT; KL-6 ≥ 500 U/mL). A bidirectional Long Short-Term Memory (LSTM) model was trained using 6-month temporal data, including weekly cough-frequency variation, monthly FeNO fluctuation, and the slope of CT fibrosis-volume change, with performance optimized through five-fold cross-validation and evaluated on the held-out validation set. The model therefore functions as a dynamic risk-stratification tool, rather than solely detecting those who already exhibit radiological worsening. Since this validation set was derived from the same single-center retrospective database, it does not constitute a genuinely independent external cohort, and these results should be considered preliminary. Public genome-wide association study data were analyzed using two-sample Mendelian randomization with the inverse-variance weighted method to assess genetic associations and pathway mediation. The fibrosis-driven phenotype accounted for 28.5% of the cohort. In this group, the 1-year PF conversion rate was 29.7% versus 9.8% in the non-fibrotic phenotypes (hazard ratios [HR] = 3.21). The early warning model achieved an AUC of 0.83 (sensitivity 85.7%; specificity 79.2%) in the validation set, identifying PF risk earlier than conventional high-resolution CT follow-up. Genetic analyses showed that variants in TGFBR2 (rs11466445) and CCL5 (rs2107538) were associated with an increased PF risk, with findings suggestive of a possible partial mediation through the TGF-β/Smad3 pathway. This study establishes a novel, data-driven framework that moves beyond static snapshots to dynamic phenotyping. The LSTM model allows early PF risk identification and suggests potential genetic mechanisms, supporting a paradigm shift towards proactive, precision management of post-COVID respiratory sequelae. Not applicable.