HealthTech NHS Bowel Cancer Screening AI — FIT Result Analysis

FIT laboratory workflow

sample arrives → barcode scan (LIMS — Laboratory Information Management System) → analyser processes sample (quantitative haemoglobin concentration — microg Hb/g faeces) → result stored in LIMS.

QC issues

• 1degraded samples (sample stored too long before processing — haemoglobin degrades, false low result),
• 2analyser drift (analyser out of calibration — systematic bias),
• 3batch errors (reagent lot issues — all samples in batch affected),
• 4outlier samples (haemoglobin concentration statistically implausible for population).

AI

anomaly detection model trained on 3 years LIMS data (840,000 results × 3 years = 2.52M training samples) → flags suspicious results for QC review.

MHRA SaMD Class I

FIT QC tool aids lab technician judgment — does not directly determine patient recall.

Class I requirements

• 1MHRA registration (DESH — Device Equipment and Software Hub),
• 2UK Responsible Person,
• 3basic safety and performance requirements (no Notified Body required for Class I).

DCB0129

quality control errors that affect FIT recall threshold application are clinical safety hazards.

Key hazard

AI incorrectly flags a valid high FIT result as a QC error → lab technician voids the result → patient who should be referred for colonoscopy is not.

Mitigation

• AI flag is advisory — lab technician must make final QC decision.
• All QC decisions logged with technician ID and reasoning.

Batch QC

each FIT analyser batch has internal controls (low control, high control, blank).

AI analyses batch control results

• 1Westgard rules (laboratory QC standard — 1-2s, 1-3s, 2-2s rules),
• 2CUSUM (Cumulative Sum — trend detection for analyser drift),
• 3peer group comparison (this analyser batch vs other analysers processing same samples — outlier batch detection).

Analyser drift detection

• moving average of last 100 results per analyser → alert if mean shifts &gt
• 5 microg from rolling baseline.

Reagent lot tracking

each reagent lot has a baseline distribution — new lot → comparison against baseline → alert if distribution shift detected.

NHS AI Lab 14 Evidence Standards applied to FIT QC tool

• 1intended purpose (quality control tool — not patient diagnosis),
• 2training data (2.52M FIT results from NHS laboratory — documented with demographics),
• 3model performance (precision, recall, F1 — flagging suspicious results vs known QC failures),
• 4external validation (200-sample validation set from different lab),
• 5bias testing (no demographic bias in QC flagging — validated by random sample of flagged results),
• 6explainability (SHAP values — which features triggered the QC flag),
• 7human oversight (lab technician always makes final decision),
• 8monitoring (monthly precision/recall check vs lab technician decisions — model drift detection).