Lightweight liquid neural networks decipher salivary metabolic fingerprinting for high-risk periodontitis screening in diabetes.

Type 2 diabetes (T2DM) and periodontitis are bidirectionally linked diseases that both involve chronic inflammation; their co‑occurrence represents a high‑risk clinical condition requiring timely identification. In this proof‑of‑concept, dual‑center feasibility study (n = 426), we evaluate a rapid, non‑invasive approach for identifying individuals with co‑occurring periodontitis and T2DM by integrating salivary metabolic fingerprints acquired by probe electrospray ionization mass spectrometry (PESI‑MS, ~0.7 min per sample) with demographic covariates (age and sex) as inputs for a lightweight liquid neural network (LNN) classifier. A total of 426 participants (H: healthy controls, n = 114; P: periodontitis, n = 209; MP: periodontitis with T2DM, n = 103) were enrolled, and the data were randomly split into training (80%) and test (20%) sets. We benchmarked the LNN against conventional classifiers (PLS‑DA, random forest, SVM) and other deep‑learning models (BiLSTM, MHA‑LSTM). While conventional models showed limited performance (AUC 0.73-0.78), deep‑learning models, which leverage the sequential structure of m/z‑ordered spectral data, achieved substantially higher accuracy. Notably, the LNN achieved the highest test accuracy (91.9%) with 100% recall for the MP group, while requiring approximately one‑third the trainable parameters of other recurrent networks. This study demonstrates the feasibility of combining PESI‑MS metabolic fingerprinting with efficient deep learning for high‑risk periodontitis screening.