KonexForge Health Insight: AI that helps read ultrasound and blood test results — a tool for doctors, not a replacement for doctors

The number of ultrasound scans and blood tests a clinic processes each day is growing faster than a doctor's available time to review every case in detail. KonexForge Health Insight is the architecture we're developing to address exactly this problem — but before describing what it does, it's worth being explicit about what it does NOT do: this is not an automated diagnostic system replacing a doctor. It's a clinical decision support tool designed to help doctors process a growing volume of data faster without missing an abnormal detail.

Why this boundary matters more than any technical feature

In line with the principles Vietnam's Ministry of Health is developing for AI in healthcare — as reported by Thanh Nien — AI plays only a supporting role for doctors and cannot fully replace clinical decision-making; every final treatment decision requires a doctor's oversight and confirmation, and the doctor bears professional responsibility for patient welfare. In the US, the equivalent legal framework is the Clinical Decision Support exemption under Section 520(o)(1)(E) of the FDCA (as amended by the 21st Century Cures Act) — CDS software is only exempt from the medical-device definition if it meets four statutory criteria, one of which requires the software to "enable a healthcare professional to independently review the basis for the recommendation" — meaning the system must be transparent about why it's suggesting something, not a black box handing down a verdict. Health Insight's entire architecture is built around exactly this constraint.

Architecture: two analysis modules, one mandatory explanation layer

**Ultrasound image analysis.** Published academic research shows real potential for computer vision in this space — for example, a systematic review on PMC found models detecting fatty liver from ultrasound images reaching roughly 90.6% accuracy with a VGG-16 architecture under research conditions, and thyroid nodule classification models reaching around 86.3% with VGG-19. These are figures from academic research on specific datasets — not a performance commitment for Health Insight, which requires its own clinical evaluation before any real deployment with a healthcare facility.

**Blood test analysis.** Similarly, published deep learning models for blood cell anomaly detection on MDPI reached a precision of 0.98 and recall of 0.99 on research datasets — combining segmentation and classification to catch abnormal blood cell patterns that are hard to spot by eye at high volume. This module of Health Insight focuses on comparing each lab value against reference thresholds and detecting abnormal combinations (not just a single outlier value) worth a doctor's closer attention.

**The explanation layer.** This is mandatory, not an optional feature: every flag from the ultrasound module comes with the specific image region the model focused on (not just a single score); every flag from the lab module comes with which value crossed a threshold, by how much, and the trend versus the previous test. The doctor always sees the basis for a suggestion, not just the final conclusion.

Router and escalation thresholds — no suggestion ever goes straight to the patient

Following the same Router and Critic Engine principles as KonexForge AI Core, every Health Insight output passes through a severity-based routing layer: results within normal range are summarized concisely for quick doctor review; results showing abnormal signs are flagged for closer attention; and most importantly — no path ever automatically sends a conclusion or explanation directly to a patient without a doctor's confirmation first, regardless of the model's confidence level. This is the jagged frontier of AI lesson applied to the highest-stakes context possible: a confident model isn't necessarily a correct model, and in healthcare, the gap between those two can directly affect a real person's health.

Health data security and continuous monitoring

Health data is among the most sensitive data categories under the zero-trust principles we covered in security for AI-integrated systems — encryption at rest and in transit, strict role-based access, and a full audit trail for every access are baseline requirements, not advanced options. In parallel, the system needs continuous model performance monitoring over time — the same principle we apply to drift monitoring in production — because case distribution, ultrasound equipment, and lab workflows differ across every healthcare facility, and a model performing well on training data doesn't automatically guarantee it performs well on a specific clinic's real-world data.

Current status: an architecture in development, seeking clinical partners for evaluation

Health Insight is currently an architecture in development at KonexForge — not yet a product that has undergone clinical evaluation or been licensed as a medical device. Before deployment at any healthcare facility, the system needs to undergo independent clinical evaluation using that facility's real data, in full compliance with current Ministry of Health regulations on clinical decision support software. This is why we're seeking partner healthcare facilities or clinics willing to jointly evaluate and refine this architecture on real data, under the supervision of their own medical teams.

Conclusion

The power of AI in healthcare doesn't come from replacing doctors — it comes from helping doctors process a growing volume of data without missing important details, provided the system is transparent about the basis for its suggestions and never skips the human-confirmation step for decisions that directly affect someone's health. This is the design principle running through Health Insight, and it's how we approach every high-stakes AI problem within the AI & ML layer. Get in touch if your healthcare facility is interested in partnering on clinical evaluation for this architecture.