Sample-efficient machine learning for dairy pathogen detection and origin traceability under multivariable conditions using gold nanoparticle colorimetry.

Abstract

Storage temperature, humidity, and time jointly influence pathogen growth in dairy products, yet most rapid detection methods are validated under single-variable conditions that do not represent real supply chain scenarios. Testing across these combined conditions sharply increases the number of samples needed, whereas dairy safety studies usually work with small data sets, which limits the use of conventional machine learning. Here we combined gold nanoparticle colorimetric sensing, a time-temperature-humidity (TTH) framework, and a sample-efficient Random Forest model for pathogen detection and geographical origin traceability in dairy products. The colorimetric assay used aptamers as recognition elements and required no complex sample preparation, nucleic acid extraction, or cold-chain reagent storage. It achieved limits of detection of 104 cfu/mL for Staphylococcus aureus, 108 cfu/mL for Salmonella typhimurium, and 107 cfu/mL for Escherichia coli, with recoveries of 96.4-108.4% in commercially sterilized fluid milk and milk powder from 4 Chinese production regions. TTH monitoring showed that at 37°C, pathogen concentrations were estimated to reach the order of magnitude of the limits in GB 29921-2021 and GB 4789.18-2024 (10² cfu/g for S. aureus and E. coli, and zero tolerance for S. typhimurium), used here as a growth-monitoring reference, within about 4 h; at 25°C this window extended to 8-9 h, while storage at 4°C kept levels low for the first ∼12 d of the 16-17-d monitoring period. To work within the small sample sizes typical of dairy safety studies, Random Forest was optimized with a target-R²-guided hyperparameter strategy. Using 2,592 measurement records from 3 independent replicates, and evaluating the models against plate-count-calibrated concentration values and the inoculated pathogen species and region of purchase as reference labels, the optimized model reached an R² of 0.993 for pathogen concentration prediction, 94.6% accuracy for multi-class pathogen classification, and 94.2% accuracy for geographical origin discrimination among Shanghai, Yunnan, Inner Mongolia, and Xinjiang samples, with areas under the curve of at least 0.97 for both classification tasks. Region-specific growth patterns in the TTH framework enabled origin identification and could support targeted recall, origin fraud screening, and source tracing at dairy distribution checkpoints. As each region was represented by a single commercial product per type and a single production batch, these results should be regarded as a proof-of-concept. This approach provides a practical screening strategy for dairy pathogen monitoring and origin traceability under multivariable supply chain conditions.