Abstract:

Background: The variability of the platelet count, due to the complex relationships between nutritional intake, environmental stressors, and individual chronophysiology, is not easily modeled by current computational methods, which typically use single-source signals and thus overlook the interacting energy-based dynamics underlying clinically relevant platelet dysregulation.

Objective: This paper constructs and confirms the existence of a multimodal machine learning system that combines diet records with high-resolution local climatic variables with body temperature time series in an innovative potential energy-based fusion mechanism to estimate absolute platelet count and identify clinically significant platelet anomalies.

Methods: We formed a multimodal group of 14,400 matched samples that were correlated with dietary characteristics and gridded indicators of microclimate (temperature, humidity, heat index, and similar variables), as well as continuous core and peripheral temperature records and laboratory measurements of platelets. Encoders, which are modality-specific, included a one-dimensional convolutional encoder on thermo-time series, a transformer-based encoder on sequential diet records, an MLP using gradient-boosted trees on climate aspects, and a convolutional image encoder when available. A PotentialEnergyFusion module maps each modality to an energy scalar and calculates data-conditioned modality weights using a negative-energy softmax. Both the fused representation and the simultaneous regression (platelet count) and classification (anomaly) heads take the fused representation as input. Training was performed using stratified 70/15/15 splits, data augmentation, standardized normalization, multi-task loss weighting, and early stopping.

Results: The proposed framework outperformed standard conjugation and attention baselines in terms of multimodal integration and interpretability, with nearly uniform variance in per-sample modality contributions, as well as predictive calibration on hold-out data.

Conclusions: Multimodal fusion that is energy-conscious records physiologically significant diet, climate, and thermoregulatory interactions, and predicts platelet counts more accurately and anomalies better. It provides an interpretable and extensible framework for haematological risk stratification and translational application.