Predict Failure
Before the X-Ray
The first study to use machine learning to predict radiation apron quality control failures from readily available data — achieving 97% precision and reducing the X-ray burden of safety programmes.
X-Ray QC Is Essential — And Exhausting
Annual X-ray-based integrity inspection of personal radiation protective equipment is the only reliable way to detect hidden defects in the attenuating layer. Visual inspection misses the vast majority of clinically significant tears. The evidence from prior studies by Kellens and colleagues (2022) and others is unambiguous: without X-ray inspection, compromised garments remain in service.
But X-ray-based QC is laborious, time-consuming, and — paradoxically — itself a source of radiation exposure for the medical physics staff conducting inspections. In large hospitals with hundreds of garments, the annual QC cycle ties up fluoroscopy time and expert resource. And because inspections occur annually, defects that develop between cycles can go undetected for months.
The question Kellens and colleagues posed in this 2024 paper is a genuinely novel one: can the outcome of an X-ray quality control check be predicted — before the X-ray — using data that is already available? If so, higher-risk garments could be flagged for more frequent inspection, and lower-risk items could be inspected less often, making the whole QC programme both more efficient and more protective.
Six Years of Quality Control Data
The study used PRPE QC records from a single large general hospital covering 2018 to 2023 — a total of 4,482 quality control checks on 1,489 unique pieces of PRPE. After excluding thyroid shields (less prone to defects due to their size) and garments with non-standard lead equivalence, the working dataset comprised 2,504 QC checks on 741 unique pieces.
Data from 2018 to 2022 formed the training set. Data from 2023 was held out as an independent test set — the most rigorous validation approach, reflecting real-world deployment conditions rather than random cross-validation. The predictors available without X-ray imaging were: brand, age, size, type, visual defects, and department.
Overall, 22.5% of QC checks in the working dataset resulted in rejection — a figure consistent with the earlier longitudinal study.
Logistic Regression vs Random Forest
Two modelling approaches were compared: logistic regression (a classical parametric statistical method) and Random Forest (RF), a non-linear, non-parametric machine learning algorithm that builds multiple decision trees and aggregates their votes.
| Metric | Logistic Regression | Random Forest |
|---|---|---|
| Accuracy | 80% | 81% |
| Sensitivity | 81% | 80% |
| Specificity | 79% | 86% |
| Precision | 95% | 97% |
| False Positive Rate | 21% | 14% |
| AUC | 87% | 87% |
The critical metric here is precision — and false positive rate. A false positive in this context means the model predicts a garment will pass QC when it would actually fail. That garment would remain in service and expose the wearer to undetected radiation through a defective shield. Minimising false positives is therefore the priority, and the Random Forest model achieves 97% precision with a false positive rate of just 14% — a meaningful improvement over logistic regression.
"A PRPE QC programme should never fully rely on a prediction model but rather use it as an additional aid to increase PRPE QC efficiency and overall PRPE quality."
Kellens et al. — Health Physics, 2024Brand Is the Dominant Predictor
Across both models, brand emerged as the single most important predictor of QC outcome — statistically significant at p < 0.001. The difference in QC pass probability between the best-performing brand (E, Protec X) and the worst-performing (A, Infinity) reached 35.1 percentage points.
In the Random Forest decision tree, brands E, D, and C trigger an immediate predicted pass with 87% probability — regardless of other variables. For brands A, B, and Other, the algorithm proceeds to examine department, age, and other factors before arriving at a prediction. Brand is, in effect, the first and most informative branch in the decision tree.
This finding has direct procurement implications. It is now possible to quantify, with statistical confidence, the degree to which the manufacturer of a garment predicts its future quality control performance. All predictors other than visual defects had a significant impact on pass probability — meaning even a brand-new garment with no visible damage can be usefully risk-stratified.
Smarter QC Without Replacing X-Ray
The authors are careful to position prediction as an aid to X-ray-based QC, not a replacement. False positives remain possible with both models — and each false positive represents a compromised garment remaining in service. The value of the model is in prioritisation: directing more frequent attention toward higher-risk garments, and potentially reducing unnecessary inspection burden for consistently high-performing brands and garment types.
- Brand is now a statistically defensible procurement criterion — a 35% difference in rejection probability should inform purchasing decisions
- Model-guided QC can stratify inspection frequency: high-risk garments inspected more often, lower-risk ones less often
- Visual inspection alone remains unreliable — the absence of visual defects was not a significant predictor of QC outcome
- Department of use and garment age are both significant predictors — intensive-use departments and older garments deserve prioritised attention
- The model should supplement, never replace, X-ray integrity analysis; false positives carry direct patient and staff safety consequences
- Future material analysis of attenuating layer composition could significantly improve model performance by replacing brand as a proxy variable
Predict Failure
Before the X-Ray
The first study to use machine learning to predict radiation apron quality control failures from readily available data — achieving 97% precision and reducing the X-ray burden of safety programmes.
X-Ray QC Is Essential — And Exhausting
Annual X-ray-based integrity inspection of personal radiation protective equipment is the only reliable way to detect hidden defects in the attenuating layer. Visual inspection misses the vast majority of clinically significant tears. The evidence from prior studies by Kellens and colleagues (2022) and others is unambiguous: without X-ray inspection, compromised garments remain in service.
But X-ray-based QC is laborious, time-consuming, and — paradoxically — itself a source of radiation exposure for the medical physics staff conducting inspections. In large hospitals with hundreds of garments, the annual QC cycle ties up fluoroscopy time and expert resource. And because inspections occur annually, defects that develop between cycles can go undetected for months.
The question Kellens and colleagues posed in this 2024 paper is a genuinely novel one: can the outcome of an X-ray quality control check be predicted — before the X-ray — using data that is already available? If so, higher-risk garments could be flagged for more frequent inspection, and lower-risk items could be inspected less often, making the whole QC programme both more efficient and more protective.
Six Years of Quality Control Data
The study used PRPE QC records from a single large general hospital covering 2018 to 2023 — a total of 4,482 quality control checks on 1,489 unique pieces of PRPE. After excluding thyroid shields (less prone to defects due to their size) and garments with non-standard lead equivalence, the working dataset comprised 2,504 QC checks on 741 unique pieces.
Data from 2018 to 2022 formed the training set. Data from 2023 was held out as an independent test set — the most rigorous validation approach, reflecting real-world deployment conditions rather than random cross-validation. The predictors available without X-ray imaging were: brand, age, size, type, visual defects, and department.
Overall, 22.5% of QC checks in the working dataset resulted in rejection — a figure consistent with the earlier longitudinal study.
Logistic Regression vs Random Forest
Two modelling approaches were compared: logistic regression (a classical parametric statistical method) and Random Forest (RF), a non-linear, non-parametric machine learning algorithm that builds multiple decision trees and aggregates their votes.
| Metric | Logistic Regression | Random Forest |
|---|---|---|
| Accuracy | 80% | 81% |
| Sensitivity | 81% | 80% |
| Specificity | 79% | 86% |
| Precision | 95% | 97% |
| False Positive Rate | 21% | 14% |
| AUC | 87% | 87% |
The critical metric here is precision — and false positive rate. A false positive in this context means the model predicts a garment will pass QC when it would actually fail. That garment would remain in service and expose the wearer to undetected radiation through a defective shield. Minimising false positives is therefore the priority, and the Random Forest model achieves 97% precision with a false positive rate of just 14% — a meaningful improvement over logistic regression.
"A PRPE QC programme should never fully rely on a prediction model but rather use it as an additional aid to increase PRPE QC efficiency and overall PRPE quality."
Kellens et al. — Health Physics, 2024Brand Is the Dominant Predictor
Across both models, brand emerged as the single most important predictor of QC outcome — statistically significant at p < 0.001. The difference in QC pass probability between the best-performing brand (E, Protec X) and the worst-performing (A, Infinity) reached 35.1 percentage points.
In the Random Forest decision tree, brands E, D, and C trigger an immediate predicted pass with 87% probability — regardless of other variables. For brands A, B, and Other, the algorithm proceeds to examine department, age, and other factors before arriving at a prediction. Brand is, in effect, the first and most informative branch in the decision tree.
This finding has direct procurement implications. It is now possible to quantify, with statistical confidence, the degree to which the manufacturer of a garment predicts its future quality control performance. All predictors other than visual defects had a significant impact on pass probability — meaning even a brand-new garment with no visible damage can be usefully risk-stratified.
Smarter QC Without Replacing X-Ray
The authors are careful to position prediction as an aid to X-ray-based QC, not a replacement. False positives remain possible with both models — and each false positive represents a compromised garment remaining in service. The value of the model is in prioritisation: directing more frequent attention toward higher-risk garments, and potentially reducing unnecessary inspection burden for consistently high-performing brands and garment types.
- Brand is now a statistically defensible procurement criterion — a 35% difference in rejection probability should inform purchasing decisions
- Model-guided QC can stratify inspection frequency: high-risk garments inspected more often, lower-risk ones less often
- Visual inspection alone remains unreliable — the absence of visual defects was not a significant predictor of QC outcome
- Department of use and garment age are both significant predictors — intensive-use departments and older garments deserve prioritised attention
- The model should supplement, never replace, X-ray integrity analysis; false positives carry direct patient and staff safety consequences
- Future material analysis of attenuating layer composition could significantly improve model performance by replacing brand as a proxy variable
Start writing here...
