Morphological signatures of bacteria‐laden drying droplets: A multimodal analysis across host fluids

Drying droplets exhibit rich morphological patterns that depend on fluid composition, yet their use in characterizing bacteria‐laden fluids remains underexplored. This paper introduces a multimodal machine learning (ML) framework that integrates numerical descriptors of texture evolution with image‐based features to examine how drying patterns vary with the concentration of Escherichia coli bacteria in three host fluids, viz., blood serum, urine, and nutrient‐rich medium. As a proof‐of‐concept, our approach exploits drying‐induced self‐assembly in bacteria‐laden host fluids as an analytical signal, providing a label‐free route to estimate bacterial concentration from microliter‐scale droplets under controlled laboratory conditions. Time‐resolved fluorescence imaging of green fluorescent protein (GFP)‐expressing bacteria further reveals distinct spatiotemporal redistribution dynamics during droplet evaporation across different fluids, providing a physical basis for the unique morphological signatures captured in the dried patterns. Rather than relying on large deep‐learning backbones, we demonstrate that optimal classification accuracy arises from a controlled balance between semantic and textural information within the image encoder. Systematic depth‐truncation experiments reveal an “optimal representational zone” where texture‐rich and biologically meaningful features combine with numerical descriptors, providing accuracy under frame‐level evaluation, with accuracy retained when tested on entirely held‐out droplets, indicating that the model captures reproducible morphological features beyond frame‐level temporal correlation. Gradient‐weighted class activation mapping (Grad‐CAM) confirms that the model attends to physically interpretable morphological cues linked to bacterial concentrations, consistent with drying droplet morphology encoding reproducible signatures across three compositionally distinct host fluids.