Why Prediction Is Not Prioritization in Healthcare AI

What I Learned from Building an AI Surveillance System for TB Control in China

Prediction alone does not determine clinical action. Operational reality does.

During a public health conference, a clinician challenged the value of our prediction models.

“These systems still don’t tell us where to deploy screening resources effectively.”

That comment stayed with me.

Later that same day, another physician working with mobile TB screening units shared a similar frustration. Some regions with seemingly high risk produced very few positive cases, while outbreaks occasionally emerged in places that had not previously appeared critical.

At that moment, I realized something important:

AI can generate risk scores.

Public health systems still need operational decisions.

They need answers to questions such as:

• Where should we screen?
• Which populations should be prioritized?
• How should limited resources be allocated?
• When should intervention begin?

Prediction alone cannot answer these questions.

From surveillance to action

Beginning in 2023, multiple regions in Zhejiang, China started deploying CAD (Computer-Aided Detection) systems to support active TB screening initiatives.

I participated in designing the surveillance and early warning module, integrating AI imaging systems with mobile screening workflows across different healthcare environments.

Initially, our approach looked similar to many digital health surveillance projects.

We focused heavily on retrospective indicators and BI-style analytics:

• positive case counts
• drug-resistant TB trends
• geographic risk distribution
• follow-up statistics
• demographic analysis

These metrics improved visibility.

But they did not necessarily improve intervention capability.

We could identify where risk had increased.

We still struggled to determine:

What should happen next?

The real gap was not prediction — it was operational prioritization

Over time, we realized that public health systems do not simply need probability outputs.

They need decision-oriented systems.

A risk score alone cannot determine:

• where to send mobile screening units
• how to allocate staff
• which regions require proactive intervention
• when epidemiological changes justify escalation

This revealed a deeper issue:

There is often a mismatch between statistical objectives and operational objectives.

AI models optimize for:

• prediction accuracy
• probability estimation
• trend detection

Public health systems optimize for:

• intervention timing
• resource allocation
• deployment prioritization
• operational response

These are not the same problem.

Rethinking the system

After several iterations, we shifted our approach.

Instead of focusing only on surveillance dashboards, we began building a more decision-oriented framework.

The system evolved to include:

Long-term risk prediction

Using:

• demographic data
• AI imaging outputs
• laboratory results
• historical epidemiological patterns

to estimate individual and regional risk over the next 12–24 months.

Geographic and mobility analysis

Integrating:

• spatial distribution
• mobile screening routes
• population movement patterns

to identify high-risk regions and vulnerable populations.

Automated early warning mechanisms

Rather than simply displaying trends, the system could proactively alert CDC teams when risk patterns changed significantly.

Adaptive deployment strategies

Helping public health teams dynamically deploy:

• screening staff
• mobile vans
• imaging resources

based on predicted operational needs.

What I ultimately learned

Many discussions around medical AI still focus primarily on:

• sensitivity
• specificity
• AUC
• model performance

These metrics matter.

But in real public health environments, AI systems compete not only with technical complexity, but also with operational reality.

A highly accurate model that cannot support real-world decision-making may still become operationally irrelevant.

Public health systems operate under constraints:

• limited staffing
• limited equipment
• uneven healthcare distribution
• aging populations
• constantly changing mobility patterns

Under these conditions, prediction alone is insufficient.

Final thought

Public health AI should not only predict risk.

It should help health systems decide:

• where to act
• when to intervene
• how to prioritize limited resources

Otherwise, even sophisticated prediction systems may remain little more than analytical dashboards — informative, but operationally disconnected from real-world action.

HealthcareAI #ClinicalWorkflow #DigitalHealth #MedicalImaging #HealthTech #Public Health