Are you worried about your FDA inspection risk?

Estimate the likely outcome of an FDA inspection

This is an app built around a LightGBM model trained on official FDA inspection data from 2009 to 2024, and tested with 2025 and 2026 data. Enter the case information, and the app will show the predicted percentages for each inspection classification alongside the evaluation matrix.

Model LightGBM multiclass F1-optimized FDA inspection model
Inputs Minimal form Only the needed ex ante fields
Output Percentages NAI, VAI, and OAI risk profile

Model background

Training data and feature engineering

Dataset

The model was trained on the FDA inspection dataset published by the FDA. Open the inspection dataset. The target is the inspection classification mapped into three classes: NAI, VAI, and OAI.

Feature engineering

The model converts country and state into regions, maps product type frequencies, derives presidential administration features, and adds inspection history counts.

Modeling approach

The app uses the LightGBM model optimized for macro F1, and keeps the exact fitted tree ensemble for inference.

Evaluation

How well does the model perform?

What the labels mean

NAI

No Action Indicated

VAI

Voluntary Action Indicated

OAI

Official Action Indicated

Each row represents the actual class, while each column represents the predicted class. The diagonal cells show correct predictions.

Context

Classification distribution over time

This chart shows how the inspection outcomes are distributed across the dataset, giving context for the class balance behind the confusion matrix.

Risk profile

Enter the case details