Before We Deploy:
Validation Frameworks
Rigorous clinical validation strategies and ethical considerations for AI diagnostic tools in healthcare — and why validation is not a box to check, but a continuous commitment.
Why Validation Is Not Optional
An AI model that achieves 95% accuracy on its training dataset and 72% accuracy on clinical deployment is not a success story. It is a patient safety incident waiting to happen. The gap between published AI performance claims and real-world clinical accuracy is one of the most consistently documented phenomena in health AI research.
The history of medical AI is littered with systems that performed brilliantly in research settings and failed at clinical deployment. A widely cited 2019 meta-analysis found that the majority of published studies used internal validation only. A 2021 systematic review found that fewer than 5% of AI diagnostic papers included prospective clinical evaluation.
In haematological AI, the stakes are acute. A morphology classification system that miscounts blasts at the 20% AML diagnostic threshold does not merely produce incorrect data — it produces incorrect diagnoses, with consequences for chemotherapy initiation, transplant eligibility, and survival.
A Five-Phase Validation Architecture
Internal Validation — Held-out Test Set
The first validation phase uses a portion of the development dataset — strictly held out from training and hyperparameter tuning — to evaluate model performance under controlled conditions. Internal validation establishes a performance ceiling, not generalisability.
External Validation — Independent Cohort
External validation tests the model on data from a different institution, population, or time period. This is the most important single validation step, testing whether the model has learned generalisable features rather than dataset-specific artefacts. Minimum recommendation: three independent external cohorts.
Subgroup and Equity Analysis
Aggregate performance metrics can conceal systematic underperformance on clinically important subgroups. Subgroup analysis should cover disease subtype, patient age, sample preparation, and — critically for LMIC deployment — geographic population and endemic disease background.
Distribution Shift and Robustness Testing
Distribution shift testing deliberately introduces controlled variations — different staining protocols, scanner models, sample handling procedures — to characterise model robustness and identify failure modes. AI systems that fail silently are more dangerous than those that fail loudly.
Prospective Clinical Evaluation
The gold standard of clinical AI validation: prospective evaluation in the intended deployment environment, measuring clinical outcomes rather than technical metrics. Pre-registered, with primary endpoints defined before data collection begins.
“The question is not whether the model is accurate on the benchmark. The question is whether it is accurate on the patient in front of you — who was not in any benchmark.”— EEHLSS Computational Haematology Team
Measuring What Actually Matters
Sensitivity (Recall)
The proportion of true positive cases correctly identified. In blast classification, sensitivity directly determines the rate of missed AML diagnoses. High sensitivity is non-negotiable in malignancy screening contexts.
Specificity
The proportion of true negative cases correctly identified. High specificity reduces false positive blast classifications that trigger unnecessary clinical workup.
AUC-ROC
Area under the receiver operating characteristic curve. AUC-ROC above 0.95 is the minimum acceptable standard for clinical deployment of haematological AI tools at EEHLSS.
Calibration
Whether the model’s confidence scores reflect true probabilities. A model claiming 90% confidence should be right 90% of the time. Miscalibrated confidence scores mislead reviewers about the reliability of specific outputs.
Agreement with Expert
Kappa coefficient measuring agreement between AI classification and expert haematopathologist review. Target kappa ≥ 0.85 for routine morphological categories; ≥ 0.75 for challenging dysplastic and blast categories.
Clinical Impact Metrics
Does AI augmentation change clinical management in a meaningful proportion of cases? Net reclassification improvement (NRI) and integrated discrimination improvement (IDI) measure the incremental clinical value added by AI.
The Context-Specific Challenge
Training Data Representativeness
An AI model trained predominantly on European or East Asian datasets may encounter morphological presentations in West African populations — sickle cell anaemia, malaria-associated haematological changes, tropical infections — that were underrepresented in its training data. The EEHLSS-WA-HemaBench project directly addresses this gap.
Analytical Platform Variation
LMIC laboratories operate diverse analyser ecosystems — Sysmex XN-series in larger hospitals, Mindray BC-6800 in mid-tier facilities, and basic manual count in resource-constrained settings. Platform-specific validation is not optional when an AI system is intended for deployment across heterogeneous analytical environments.
| Validation Requirement | ISO 15189:2022 | MLSCN (Nigeria) | CBAHI (KSA) | EEHLSS Standard |
|---|---|---|---|---|
| Internal validation | ✓ Required | ✓ Required | ✓ Required | ✓ Mandatory |
| External validation | ◑ Recommended | ◑ Partial | ◑ Partial | ✓ Mandatory |
| West African cohort | — Not specified | ◑ Implicit | — N/A | ✓ Mandatory |
| Prospective evaluation | ◑ Recommended | — Not specified | ◑ Partial | ✓ Mandatory |
| Continuous monitoring | ✓ Required | ✓ Required | ✓ Required | ✓ Mandatory |
All AI components in MedLabAI-LIS and HemaPath undergo the full five-phase validation framework before clinical deployment. EEHLSS exceeds ISO 15189, MLSCN, and CBAHI minimum requirements by mandating external validation on geographically and demographically diverse cohorts — including a dedicated West African validation dataset.
The Ethics of Deploying Before We Know
Equity and Algorithmic Fairness
If an AI model performs significantly better for patient groups well-represented in its training data, deployment in underrepresented populations constitutes a systematic equity failure. Algorithmic fairness auditing is a mandatory component of pre-deployment ethics review.
Informed Deployment and Scope Limitations
Clinicians who rely on AI diagnostic tools have a right to know the boundaries of validated performance. At EEHLSS, every HemaPath AI tool ships with a Performance Summary Card: a structured document specifying the training cohort, external validation cohorts, performance metrics by subgroup, known limitations, and the deployment scope within which the published performance applies.
Leave a Reply