跪拜 Guibai
← Back to the summary

Medical AI's Real Gatekeeper Isn't Model Accuracy—It's the Hospital IT Security Review

When Healthcare Meets AI Agents: Why Coco Keeps Collaboration Inside the Hospital Intranet

Introduction

In April 2026, the National Health Commission launched the "Three-Year Action Plan for Improving Medical Quality in Primary Healthcare Institutions," promoting the downward allocation of quality medical resources and intelligent upgrades. In March of the same year, the "2026 National Medical Quality and Safety Improvement Goals" were released, embedding goals like "improving the substantive quality of medical records" and "reducing the incidence of medical errors" into the system. The door to medical AI has been opened—but to truly walk through it, one question must be answered first: where does patient data go?

The Healthcare AI Paradox: Most Urgent Demand, Most Difficult Implementation

Among all industries, healthcare's demand for AI may be the most genuine.

Primary care physicians lack expert support when facing difficult cases. The medical records department of a tertiary hospital manually reviews thousands of records every month. Clinical research teams screen subjects from massive volumes of medical records, measuring efficiency in "weeks." Preparations for multidisciplinary consultations—organizing patient histories, test results, and imaging reports—consume a significant amount of residents' time.

But healthcare is also the most difficult industry for AI implementation. Not because of a lack of technology, but because the constraints are too rigid:

Patient privacy is an untouchable red line. The Personal Information Protection Law, the Data Security Law, and the Cybersecurity Management Measures for Medical and Health Institutions—if medical data leaves the hospital intranet, the trigger is not an experience problem, but a legal one. No matter how powerful an AI tool is, if it requires uploading medical records to the cloud, the hospital IT department's answer is just two words: not possible.

Reliability requirements are a matter of life and death. An error in power dispatch can be rolled back. An error in medical decision support—missing a contraindication, misjudging a drug interaction—the cost is not "degraded user experience," but a clinical accident.

Knowledge updates cannot be interrupted. Clinical guidelines are updated annually, drug inserts are revised at any time, and diagnostic and treatment standards iterate continuously. If AI relies on a model trained a year ago, the advice it gives may already be outdated or even dangerous.

Primary care needs AI more, but its IT foundation is the weakest. Township health centers and community health centers—the places where AI can best compensate for the shortage of specialists—often have the most unstable network conditions and the most limited IT operation and maintenance capabilities. Solutions that rely on the cloud and complex deployment are "installed but unusable" at the primary care level.

These constraints stack together, requiring medical AI to meet a set of seemingly contradictory conditions: strong capability, but data cannot leave; continuous updates, but no reliance on remote services; simple enough, but without sacrificing reliability.

What is Coco: An AI Partner That Keeps Both Data and Knowledge Local

Before entering the medical scenario, let's clarify its identity.

Sinoaus Coco is not a "medical AI product." It does not understand diagnosis out of the box, does not preset medical knowledge, and does not replace any doctor's judgment.

Sinoaus Coco is a general-purpose AI collaboration partner—running on a local machine, data does not leave the execution environment, the underlying model is replaceable, and team knowledge can be solidified. Its architectural design revolves around a core premise: the collaborative relationship is private and should not be defined by any third-party server, network condition, or commercial agreement.

Several key features:

Local-first. Conversation context, case summaries, review standards, accumulated experience—all data that constitutes the "partner's memory" is stored on the local disk. It does not pass through external servers and does not depend on the availability of cloud services. For the healthcare industry, this choice is not about "better privacy protection," but the entry ticket for "passing the IT department's security review."

Model-agnostic. The underlying inference model can be replaced—a domestic model can be deployed on the intranet, or a cloud API can be used in non-confidential scenarios. Switching does not affect the collaborative relationship. The diagnostic and treatment norms, quality control standards, and historical cases that the partner remembers are not bound to any single model.

Skill system. Teams can solidify knowledge and processes into Skills—a medical record quality control checklist, drug interaction verification rules, clinical research inclusion/exclusion criteria templates. Coco automatically loads them in relevant scenarios, eliminating the need for manual reminders each time.

Long-term memory. No context length limit. Difficult case discussions, consultation opinions, and medication experience feedback accumulated by a department—the partner remembers them continuously, without loss.

Scenario 1: A Primary Care Clinic, Facing an Uncertain Case

A doctor at a township health center faces hundreds of diseases every day but has no senior expert nearby to consult at any time. Encountering a patient with atypical symptoms and borderline test results—"Should this patient be referred?"—the pressure of this judgment is hard for a tertiary hospital doctor to understand.

Referring too early inconveniences the patient and squeezes higher-level hospital resources. Referring too late delays treatment and may lead to an accident.

If the primary care doctor had an AI partner on their computer—not remotely connected to a cloud service (primary care networks often lag even for video consultations), but running directly locally—what could it do?

It could do three things: First, after inputting the patient's chief complaint, physical examination results, and existing test data, quickly search the locally stored latest diagnostic and treatment guidelines for matching content, marking "red flag signs" that need vigilance. Second, check against the drug insert database for any interactions between prescribed medications and conflicts with the patient's allergy history or kidney function indicators. Third, if the system marks indications requiring referral, automatically generate a structured referral summary—chief complaint, physical examination, tests performed, preliminary judgment, referral reason—without the doctor needing to manually compile it.

Throughout the entire process, patient data never left this computer. The inference model deployed on the intranet completed all calculations, and medical record information did not pass through any external service.

Image

Scenario 2: The Medical Records Department, 3,000 Records for Quality Control Monthly

The medical records department of a tertiary hospital faces quality control reviews of three to five thousand discharge records every month. Is the primary diagnosis selection correct? Do the surgical codes correspond? Are the progress notes complete? Are there missing signatures on informed consent forms?

The traditional approach is for quality control officers to review them one by one, comparing them item by item based on experience and the quality control manual. Efficiency depends on the officer's focus and experience—a skilled officer can review at most 40-50 records per day. The ideal review cycle for 3,000 records is one and a half months. In reality, when manpower is tight, many records are "spot-checked" rather than "fully reviewed."

Sinoaus Coco's approach is to solidify the quality control manual into review Skills: primary diagnosis selection rules, surgical operation code mapping, mandatory field completeness checks, and time-bound compliance verification. After the quality control officer inputs the medical record summary, the partner verifies item by item, marking non-compliant items—"Primary diagnosis does not match surgical code," "Post-operative progress note exceeds the specified time limit from surgery date," "Blood transfusion informed consent form lacks patient signature."

The quality control officer's energy shifts from "checking every item in every medical record for omissions" to "reviewing the anomalies flagged by the partner." With unchanged manpower, coverage changes from spot-checking to full review. Moreover, Skills can be updated as quality control standards update—next month, if the Health Commission issues new medical record management standards, simply updating the Skill file automatically synchronizes the quality control standards.

Scenario 3: Multidisciplinary Consultation, Preparation Time from Two Hours to Ten Minutes

Tumor Multidisciplinary Team (MDT) consultations are a daily operation in large hospitals—experts from medical oncology, surgery, radiotherapy, imaging, and pathology gather to formulate comprehensive treatment plans for complex cases.

The consultation itself usually lasts only 15-20 minutes. But the preparation work residents do for each consultation—compiling the patient's entire medical history from initial diagnosis, organizing test results from various departments, retrieving key screenshots from PACS images, and sorting through previous treatment plans and efficacy evaluations—takes an average of two hours.

The pain point in this process is not "difficulty," but "fragmentation." Information is scattered across four systems: HIS, LIS, PACS, and Electronic Medical Records, with inconsistent formats and unclear timelines. Residents need to jump, copy, organize, and format like detectives between different systems.

Sinoaus Coco's role in this scenario is not to replace the resident's clinical judgment. It is to do what it can already do—read structured data, arrange it by timeline, categorize it by department. The resident describes the basic situation of the case in natural language, and Coco extracts key information from locally stored medical record summaries and historical examination records, generating a draft consultation summary arranged by timeline and annotated by department. The resident only needs to verify and supplement—from two hours to ten minutes.

Image

What's important: the entire process is completed within the hospital intranet. All of the patient's medical history information did not pass through the internet and was not uploaded to any third-party server.

The Core Contradiction in Medical AI is Not "Insufficient Capability," but "Insufficient Trust"

The healthcare industry's attitude towards AI has always been ambivalent: on one hand, acknowledging its potential; on the other, refusing to entrust any critical link to it. This contradiction is not because doctors are conservative—it's because the trust threshold for AI in medical scenarios is higher than in any other industry.

The foundation of trust, broken down, consists of three things:

Data trust. Patient information has never left the hospital's controllable range. It's not "we promise encrypted transmission," but "transmission simply did not happen."

Capability trust. The advice given by AI is based on the latest versions of clinical guidelines, drug inserts, and diagnostic and treatment norms—not based on what it "saw" during training a year ago. Moreover, the source of this basis is transparent and traceable.

Reliability trust. Network down? AI can still work. Supplier goes bankrupt? The system can still run. Model replaced? Knowledge and processes are unaffected.

Sinoaus Coco's architecture happens to make choices in these three dimensions: data stored locally (data trust), knowledge continuously updated through Skills (capability trust), model replaceable and running locally (reliability trust).

This is not a feature "customized" for the healthcare industry. It is the natural adaptation of a local-first architecture across different industries.

Conclusion

In 2026, the National Health Commission drew a clear starting line with the "Three-Year Action Plan": primary care quality must improve, and intelligence must be pushed downward. But what truly allows AI to enter hospitals, departments, and township health centers is not "whose model scores higher on benchmarks," but "who can pass the IT department's security review."

Coco's answer has never changed: the partner lives locally, data stays on the intranet, and knowledge solidifies into the team's permanent asset.

The breakthrough in medical AI does not lie in another 10 percentage point increase in model performance. It lies in allowing the IT department to check the box on the security review form.

References:

Image Coco Official Website: https://coco.sinoaus.net

Comments

Top 1 from juejin.cn, machine-translated. The original thread is authoritative.

古马

The biggest upside of medical AI is that you can trust its medical ethics.