MICO-PMInet: An Interpretable Multi-Organ Incomplete-Input Cooperative Learning Framework Based on ATR-FTIR Spectra for Robust Postmortem Interval Estimation

Abstract

Accurate estimation of the postmortem interval (PMI) is a central challenge in forensic practice, particularly when faced with organ absence, tissue destruction, or restricted sampling. Current approaches largely rely on single-source specimens or complete-input assumptions. To address these challenges, this study developed MICO-PMInet, an interpretable multi-organ incomplete-input cooperative learning framework based on attenuated total reflectance-Fourier transform infrared (ATR-FTIR) spectra from eight organs. MICO-PMInet simulates real forensic reasoning to implement PMI inference via specimen screening, single-organ feature extraction and multi-organ fusion. Four key components, including organ feature encoding, inter-organ information interaction, multi-organ prediction aggregation, and incomplete-input training strategy, were systematically designed and compared to enable robust PMI prediction under variable organ-combination and partial-missing-input conditions. Systematic comparison showed that the OSB-WMHA-AWA-MORM configuration achieved the best balance among prediction accuracy, missing-input robustness, and organ heterogeneity modeling. It reaches R2 = 0.90, RMSE = 2.56 h and MAE = 1.62 h on test data, with better performance using all eight organs. Satisfactory prediction is still guaranteed using merely one organ, confirming its strong adaptability to missing samples. SHAP results reveal liver, kidney and heart dominate prediction, and core spectral bands correspond to typical biomolecular vibrations. Collectively, MICO-PMInet provides a proof-of-concept framework for PMI estimation under variable organ availability in a controlled rat setting, while its transferability to human forensic applications requires further validation.