Navigating The Pitfalls of AI Development and Implementation: The NUH Scoliosis AI Project Custom Case Solution & Analysis

Case Evidence Brief: NUH Scoliosis AI Project

1. Financial Metrics

• Screening Volume: The School Health Service in Singapore screens over 100,000 students annually for scoliosis.
• Resource Cost: Manual Cobb angle measurement requires 3 to 5 minutes per radiograph by a trained specialist.
• Development Investment: Internal funding from National University Health System (NUHS) for AI development and clinical validation.
• Market Potential: Global scoliosis management market valued at approximately 3 billion dollars, with a growth rate of 4 percent.

2. Operational Facts

• Current Process: Manual identification of the most tilted vertebrae (end-vertebrae) followed by geometric calculation of the Cobb angle.
• Technical Specification: AI model utilizes deep learning architectures, specifically convolutional neural networks trained on historical X-ray data from NUH.
• Dataset Size: Initial training utilized over 5,000 labeled spinal radiographs.
• Performance Data: AI achieves a mean absolute error of less than 3 degrees compared to senior radiologists, which is within the acceptable clinical margin of 5 degrees.
• Geography: Project based in Singapore, operating under the Health Sciences Authority (HSA) regulatory framework.

3. Stakeholder Positions

• Dr. Kevin Lim: Lead Orthopedic Surgeon; advocate for clinical utility but emphasizes that AI must assist rather than replace human judgment.
• AI Development Team: Focused on model accuracy and reducing false negatives to ensure no cases of severe scoliosis are missed.
• Hospital Administration: Prioritizes operational efficiency and the reduction of wait times for specialist consultations.
• Regulatory Bodies (HSA): Require strict validation of safety, efficacy, and data privacy before granting Class B medical device status.

4. Information Gaps

• Infrastructure Costs: The case does not specify the exact cost of cloud computing resources or on-premise hardware required to run the AI at scale.
• Liability Framework: Specific legal protocols for misdiagnosis by the AI remain undefined.
• Long-term Maintenance: Budget for model retraining to prevent performance decay over time is not explicitly stated.

Strategic Analysis: Transitioning from Pilot to Clinical Standard

1. Core Strategic Question

• How can NUH integrate the scoliosis AI into the clinical workflow to maximize diagnostic efficiency without compromising patient safety or incurring prohibitive regulatory liability?

2. Structural Analysis

Value Chain Analysis: The diagnostic value chain is currently bottle-necked at the interpretation stage. Manual measurement creates a lag between image acquisition and treatment planning. The AI shifts value from labor-intensive measurement to high-level clinical decision-making. However, the value is lost if the surgeon spends equal time verifying AI outputs.

Jobs-to-be-Done: The surgeon needs to determine if a patient requires bracing, surgery, or observation. The AI job is not just to measure an angle but to provide a reliable filter that removes healthy patients from the specialist queue, allowing focus on complex cases.

3. Strategic Options

4. Preliminary Recommendation

NUH should pursue the Augmented Decision Support model for the next 24 months. This path allows the team to collect real-world performance data while maintaining the surgeon as the final checkpoint. This strategy minimizes liability during the initial rollout and builds the necessary trust among clinical staff before attempting full automation. The primary requirement is a seamless integration into the existing Radiology Information System (RIS).

Operations and Implementation Plan

1. Critical Path

• Phase 1 (Months 1-3): Finalize integration with the Picture Archiving and Communication System (PACS). This is the technical foundation for all subsequent steps.
• Phase 2 (Months 4-6): Conduct a prospective clinical trial with 1,000 patients to compare AI-assisted measurements against a three-surgeon consensus.
• Phase 3 (Months 7-9): Submit the validation data to the Health Sciences Authority for Class B medical device certification.
• Phase 4 (Month 10+): Full deployment in the NUH pediatric orthopedic clinic.

2. Key Constraints

• Technical Debt: The AI must handle various X-ray formats and qualities from different machines across the health cluster.
• Surgeon Adoption: Resistance may occur if the AI interface adds even 30 seconds of extra clicking to the workflow.
• Data Privacy: Compliance with the Personal Data Protection Act (PDPA) in Singapore is non-negotiable, requiring strict de-identification protocols.

3. Risk-Adjusted Implementation Strategy

To mitigate the risk of model drift, the team will implement a continuous monitoring dashboard. If the variance between AI and surgeon measurements exceeds 5 degrees in more than 10 percent of cases over a month, the system will revert to manual mode for a mandatory audit. Contingency planning includes maintaining a pool of locum radiologists to handle surges if the AI system requires downtime for recalibration.

Executive Review and BLUF

1. BLUF

The NUH Scoliosis AI project is a technically sound initiative that addresses a critical bottleneck in pediatric spinal care. To succeed, the project must shift focus from model accuracy to workflow integration. The recommended path is an augmented intelligence model where the AI serves as a specialist assistant. This approach balances the need for efficiency with the necessity of clinical oversight and regulatory compliance. Immediate priority must be given to PACS integration and securing HSA certification. Execution success depends on minimizing the friction of the user interface for surgeons.

2. Dangerous Assumption

The single most consequential premise is that high accuracy in a retrospective dataset will translate to clinical utility. This ignores the reality of varied X-ray positioning and image noise in a high-volume screening environment, which may degrade AI performance in ways not seen during training.

3. Unaddressed Risks

• Liability Gap: The analysis does not fully address who bears the legal burden if the AI fails to flag a progressive curve that later requires surgery. Consequence: Potential for significant litigation and reputational damage to NUH.
• Technical Obsolescence: Rapid changes in imaging hardware or the emergence of low-dose radiation alternatives could render the current model obsolete. Probability: Moderate.

4. Unconsidered Alternative

The team has not evaluated the potential of a mobile-based screening tool for parents. By utilizing the camera on a smartphone to detect trunk asymmetry (the Adam Forward Bend test), the hospital could filter the patient population before they even enter the radiological workflow. This would address the volume problem at the source rather than at the diagnostic stage.

5. Verdict

APPROVED FOR LEADERSHIP REVIEW