Development of a Decision Support System for the Assessment and Prediction of Food Contamination Risk
Neotron is an ACCREDIA-accredited analytical laboratory based in Modena, part of the Cotecna Group, operating in testing, inspection, and certification across food, feed, pharmaceutical, and environmental sectors. With 550+ staff and a high-throughput instrument fleet (30+ LC-MS/MS, 20+ GC-MS/MS), it accumulates a large proprietary occurrence dataset whose analytical potential goes well beyond individual pass/fail judgements.
The internship is hosted within the LMIB Division, where analytical data are processed and integrated with regulatory frameworks to support evidence-based decision-making in food safety monitoring.
From raw LIMS extractions to structured datasets, statistical models, and an operational Decision Support System for food contamination risk assessment.
The project transforms a routine compliance archive into an operational risk intelligence tool, combining statistical inference, machine learning, and biological characterisation of priority contaminants. The pipeline spans multiple food matrix categories and contaminant classes across the six regulatory sections of Regulation (EU) 2023/915.
Historical analytical records from Neotron's LIMS, extracted as structured workbooks via SQL queries. Regulatory limits are derived dynamically at query time, ensuring the system stays aligned as the regulation is updated.
Regulation (EU) 2023/915 organises contaminants into six thematic sections: mycotoxins, plant toxins, heavy metals, POPs/PFAS, process contaminants, nitrates/perchlorates. Its living nature, updated continuously as EFSA adopts new opinions, motivates a dynamic computational approach.
Modular, reproducible pipeline in Python 3.10 across four Jupyter notebooks, each controlled by a single section variable for full reproducibility across the six regulatory sections.
SQL extraction from the LIMS. Raw result field parsing: administrative strings dropped, not-detected values mapped to 0, left-censored entries dropped. Descriptive statistics, box plots, NC rate bar charts, contaminant by month heatmaps. Normality testing via Shapiro-Wilk or Anderson-Darling. Temporal trend analysis via Mann-Kendall on deseasonalised monthly NC rates with Sen's slope and multiple testing correction. PCA and hierarchical clustering of food matrices.
Unit of analysis: contaminant / sub-matrix pair. Fisher's exact test with multiple testing correction identifies pairs with significantly elevated NC rates. Wilson confidence intervals quantify uncertainty around each estimate. DSS risk thresholds calibrated empirically from the observed NC rate distribution, producing a HIGH / MEDIUM / LOW classification per pair.
Binary classification for pairs with insufficient observational history for direct statistical inference. Leakage-free features derived from dataset topology only. Four models (Random Forest, XGBoost, LightGBM, CatBoost) with native class weighting, stratified 5-fold cross-validation, and PR-AUC as primary metric. SHAP values for global feature importance and individual-level explanations. Final DSS score combines statistical and predicted signals.
High-risk contaminants mapped to chemical-gene interactions via the Comparative Toxicogenomics Database, queried via REST API for human and mouse targets with direct experimental evidence only. GO and KEGG pathway enrichment analysis with multiple testing correction. Tripartite contaminant-gene-pathway network built with NetworkX, visualised interactively via pyvis. Hub genes connected to multiple contaminants flagged as priority signals for RASFF alert communication.
Outputs: cleaned datasets, section-level EDA reports, risk-scored contaminant-matrix catalogue, ML model cards with SHAP explanations, interactive biological network, and a Power BI dashboard for operational monitoring.