Temporal Deep Learning for Predicting Periodontitis Progression Using Longitudinal Gingival Crevicular Fluid Protein Profiles

Background: Conventional clinical indicators of periodontitis progression detect disease after irreversible tissue destruction has occurred. Molecular biomarkers in gingival crevicular fluid (GCF) offer potential for earlier detection, but existing analytical approaches rely on cross-sectional snapshots that fail to capture the temporal dynamics of disease evolution. Aim: To develop and validate a temporal deep learning framework leveraging longitudinal GCF protein profiles for (1) regression-based prediction of clinical attachment level (CAL) and probing depth (PD) changes, (2) current-visit classification of periodontitis progression, (3) next-visit prediction of progression with a 2-month clinical lead time, and (4) identification of the most informative biomarkers through systematic multi-method feature importance analysis. Materials and Methods: This study utilized longitudinal GCF data from a prospective cohort of 413 participants (501 periodontal sites, 3,792 time-series observations) with 64 protein biomarkers measured at 2-month intervals over 12 months. A compact encoder-gated recurrent unit (GRU)-decoder architecture was developed through systematic experimentation across four phases, benchmarking temporal deep learning against cross-sectional machine learning baselines. Task-specific decoders addressed continuous regression (CAL and PD prediction) and binary classification (progression detection). Model development and reporting followed the TRIPOD+AI guidelines. Results: The temporal GRU achieved 47.7% CAL mean absolute error (MAE) reduction (1.139 to 0.596 mm) and 41.0% PD MAE reduction (0.902 to 0.532 mm) over linear regression baselines through the systematic model development progression. For binary classification, the model achieved AUC-ROC of 0.886 for current-visit classification and 0.867 for next-visit prediction with a 2-month lead time. Per-visit analysis revealed progressive improvement in both regression and classification accuracy as longitudinal data accumulated. Cross-method feature importance analysis identified Periostin, VEGF, MMP-2, IL-1RA, and MCP-4 as core predictive biomarkers, with divergent profiles between diagnostic and prognostic tasks suggesting distinct molecular signatures for concurrent versus incipient progression. Conclusions: Temporal deep learning applied to longitudinal GCF protein profiles enables both accurate regression prediction of clinical parameters and reliable classification of progression status, including 2-month-ahead forecasting suitable for clinical intervention planning. The compact architecture and non-invasive sampling approach make this framework suitable for integration into point-of-care periodontal monitoring workflows.