[The content of this article has been produced by our advertising partner.]
The Radiology Department fields a steady flow of enquiries from many patients at the same time, arriving at all hours of the day and night, often stretching over days or months as patients consider their options or return after consulting their referring doctor. Each time a conversation resumes, staff have to pick up where it left off. The underlying work is complex too: 1,000+ distinct examination items, each with pricing logic that depends on modality, contrast, anatomy, and applicable packages, and referral letters that arrive as photos with handwritten notes and mixed-language text. Doing all of this accurately, quickly, and around the clock is genuinely demanding, and simply hiring more staff is not a sustainable solution.
Yet the patient experience is straightforward. A patient sends a referral letter via a dedicated WhatsApp channel to the Department. Within minutes, they receive a reply in their own language. The correct examination is identified. A quotation is matched to the Department's pricing. The same experience runs at any hour, in either Chinese or English.
AI now handles the routine intake and quotation work, with staff kept in the loop for review. When a patient confirms they want to proceed, their appointment request is logged and surfaced to staff for checking and action. Every conversation is also categorised and tagged automatically, so when staff come on shift, they see a clear, organised overview of items requiring attention. They can pick up where the AI left off, focusing their attention on the patients and clinical decisions.
Why the Standard AI Playbook Fails in Healthcare
The usual way to introduce AI into a workflow is straightforward: call a hosted model API, send the message, wait for a response. For a radiology department handling thousands of patient messages each month, that pattern doesn't work, for two main reasons.
The first reason is the nature of the data involved. Referral letters routinely contain patients' names, dates of birth, and other personal data. From a data protection and confidentiality standpoint, this information cannot be sent to external services in raw form. Equally important, the de-identification step cannot be outsourced, because the goal is to prevent third parties from ever accessing raw patient information.
The second reason is the nature of the documents themselves. Real referral letters include handwritten annotations, physician shorthand, and clinical formats that vary across referring institutions. General-purpose models do not reliably handle this kind of material out of the box.
Bringing the Model Inside the Walls The architecture splits the AI work along privacy lines.
The components that handle raw patient data, including document processing and the de-identification layer, run on Vascue's own infrastructure within AWS's ap-east-1 (Hong Kong) region. Identifying details are masked at the boundary before any content is passed onward.
AWS Hong Kong Solution Architects worked directly with Vascue's engineering team on the deployment design, covering network isolation, access controls, compute boundaries and scaling. Local GPU-accelerated compute in the Hong Kong region was a hard prerequisite for the architecture to work.
Once direct identifiers have been removed, the system further processes the content so that only the data strictly necessary for the downstream AI tasks is passed to subsequent processing. Broader re-identification risks are mitigated through Vascue's technical and security controls, including data minimisation, access controls, audit logging, and contractual restrictions on downstream processing.
The components are also tuned rather than generic. Off-the-shelf models do not survive contact with real-world medical documents, and a misread annotation is not a UI glitch. It changes how a patient enquiry is understood. Vascue's pipeline is tuned for the department's specific clinical formats and conventions. The de-identification layer is monitored, logged, and reviewed regularly as part of Vascue's ISO 27001 controls, with humans in the loop for ongoing review.
Both layers are independently audited. AWS's infrastructure carries ISO 27001, 27017, 27018, SOC 2, and CSA STAR certifications, and Vascue maintains ISO 27001.
What Becomes Possible
In production, the system has served 6,500+ patients, processed 110,000+ messages, and supported 3,300+ successful bookings. Text-based enquiries typically receive a response within 20-30 seconds. For image-based referrals, where the system extracts content from the document, removes identifying details, matches the right examination against the Department's pricing, and generates a quotation reply, the end-to-end turnaround is typically around 60 seconds. WhatsApp support runs 24/7, matching the hospital's clinical services, and handles an average of 46 after-hours enquiries per day. This is patient demand that would otherwise queue up overnight or go unanswered, now resolved at the time the patient is actually thinking about it. Staff rate 99.2% of AI responses as correct, with humans in the loop for ongoing review.
The Radiology Department's deployment shows what the architecture can do in production. The same pattern applies across clinical contexts: sensitive processing remains inside Vascue's environment, identifiers redacted at the boundary, staff in the loop for clinical decisions.
Bringing AI into healthcare at scale takes domain expertise, careful systems integration, and ongoing operational discipline. This deployment is one working example of what that looks like in a regulated environment.