跪拜 Guibai
← Back to the summary

MedDoc AI: One Workbench for Clinical Writing, Collaboration, and Evidence-Based AI

MedDoc AI: Putting Clinical Writing, Collaboration, Quality Control, and Knowledge Q&A into One Workbench

Author: 徐小夕

Over the past period, we have been researching and refining AI use cases in medical informatization. After talking with many clinicians, quality control staff, and information department colleagues, the clearest takeaway is: a huge amount of their energy is consumed by trivial, record-related chores.

To write a progress note, you have to go through medical history, look up lab results, find imaging, and switch between four or five systems. When multiple people collaborate on one record, versions pile up and get messy, and nobody can say who changed what. Quality control is mostly retrospective; by the time problems are discovered, archiving is often close, leaving no time for proper corrections. Many AI-powered medical record tools only focus on generation without supporting evidence, so the content can't be traced, and clinicians don't dare actually use it.

Ultimately, medical records are not ordinary office documents. They are both a record of the treatment process and have legal evidentiary value, so stability and accuracy always matter more than speed.

Based on this judgment, we built a clinical documentation workbench called MedDoc AI on top of the JitWord collaborative AI document foundation.

The original intention of this project is simple: not chasing flashy concepts, but focusing on integrating clinical writing, multi-person collaboration, medical record quality control, and medical knowledge Q&A into one workspace, reducing system switching and repetitive work, ensuring every modification has a basis, and every document is traceable.

This article starts with real clinical pain points, dissects the system's design philosophy, core capabilities, and practical value, and also clarifies suitable scenarios and boundaries.

Common Pain Points in Clinical Medical Record Management

After visiting institutions of various sizes, we found that many problems are shared across the board, regardless of level or department—they are industry-wide.

Heavy writing burden, energy drained. Doctors spend a large portion of their work time on medical record writing, formatting adjustments, and copying excerpts. Past history, test results, and medication records are scattered across different systems; writing a record requires switching pages repeatedly to copy and paste. The energy that should go into clinical judgment and treatment planning is consumed by mechanical paperwork.

Disordered multi-person collaboration, unclear versions. A single medical record often goes through multiple rounds of edits by the attending doctor, senior physicians, and quality control staff. In the traditional mode, files are passed back and forth, version naming depends on personal habits, and it is hard to trace who changed what and when. Comments and suggestions are scattered in chat tools, disconnected from the document itself, so rectification and checking require repeated comparisons, which is inefficient.

Retrospective quality control, high rectification costs. Most hospitals still use a post-hoc sampling approach for medical record QC. By the time writing defects or deadline issues are discovered, archiving is often near, corrections are rushed, and errors recur. Time limit monitoring, norm checks, and content completeness cannot be reminded during the writing process; they all rely on manual post-hoc inspection, which is inefficient and prone to misses.

AI lacks evidence, so it's not reassuring. Many AI medical record tools only pursue generation speed, and output has no source support, easily leading to factual errors. Given the seriousness of medical documents, AI-generated content without an evidence chain not only harms record quality but also carries compliance risks. Clinicians prefer to write a bit more themselves than to use unsubstantiated content.

Information too scattered, hard to grasp the full picture. Admission records, progress notes, test reports, and imaging are scattered across different business systems. To understand a patient's condition comprehensively, doctors have to open each portal to query. Information is fragmented, and key indicators or historical anomalies are easily missed, affecting judgment efficiency and creating clinical risks.

These real pain points made us decide on the product direction: not a single-point tool, but a complete workbench covering writing, collaboration, quality control, and knowledge assistance, so clinicians switch less, run around less, and redo less.

Core Capability Design: Built Around the Real Clinical Workflow

We did not pile up features; we went through the complete clinical document workflow and solved each stage's pain points. Below we explain each capability, what scenario it corresponds to, and what problem it solves.

Doctor Workstation: To-dos and Deadlines at a Glance

The most annoying thing in clinical work is hunting for things to do—which records are pending, which are near timeout, which are waiting for review. They are scattered across different pages, easy to miss.

We aggregated all document tasks a doctor handles daily onto the workstation home page. After login, the records a doctor is responsible for are clear: which are drafts, which are submitted for review, which are nearing the writing deadline. The first progress note overdue warning and archiving reminders are pushed proactively, so you don't need to remember deadlines and check every page.

For doctors, the most direct feeling is peace of mind: no need to worry about various time nodes, the system monitors progress for you; to-dos are centralized, so you don't have to jump around. This seemingly small aggregation actually saves a lot of fragmented switching time.

Collaborative Medical Record Editing: Multi-person Editing with Full Traceability

For the scenario of multiple people editing a record, we fully integrated JitWord's online editing capability and added medical-specific permissions and workflow controls.

On the same record, the attending doctor writes the draft, the senior physician adds online comments and edits, and the QC staff marks defects simultaneously. Everyone works on the same document without passing files back and forth. Every save generates a version record. Who changed which part at what time can be traced back at any time. Comments are written directly in document annotations, linked to specific paragraphs, so rectification is easy to align.

Compared to traditional offline file transfer, the improvement in collaboration efficiency is immediate. More importantly, all changes are traceable, meeting the traceability requirements of medical documents; if a problem arises, it can be located to a specific step and responsible person.

Patient 360 Profile: Full Data Connected on a Timeline

Many doctors have this experience: understanding a patient's full picture takes more time than writing the record itself, because you have to run several systems to look up data.

In MedDoc AI, opening a patient's profile shows a complete timeline from admission to the present: basic info, all past medical records, test reports, and vital sign records, all connected in chronological order. Authorized data is displayed collectively, so no need to jump between multiple systems.

The value is more than saving time. With continuous information, doctors get a more comprehensive grasp of the condition, and can visually see historical anomalies and treatment changes without missing key clues. Of course, all display strictly follows permission rules; content you are not authorized to see is blocked at the backend, not just hidden from the UI.

Real-time Quality Control: Moving Checks Forward, Fixing While Writing

The biggest problem with traditional QC is lag: by the time defects are found, little time remains before archiving, corrections are rushed, and often not thorough.

We embed QC rules into the writing process. While a doctor writes, the system checks in real time—for example, missing required fields, upcoming writing deadlines, and format compliance. Problems are flagged on the spot, so you fix while writing, without waiting for a final batch rework.

The QC side also has a dedicated workbench where all defects are displayed by category and severity: fatal, important, and warning. Problems can be dispatched online for correction, progress tracked throughout, and after correction, it automatically enters re-review. From discovery to closure, the whole process is online, no more offline communication or paper registration.

For QC staff, no more flipping through records one by one; efficiency and coverage improve. For clinicians, problems are caught early and fixed, avoiding a pile-up before archiving, which improves the experience.

Medical Knowledge and Evidence Q&A: Every Answer Has a Source

In medical AI, trust is always the first priority. So from the start, we did not build gimmick features like "fully automatic record writing"; we took the evidence-based AI route.

Whether a doctor asks about record-writing standards, clinical knowledge, or searches data based on the current record, the AI first retrieves evidence within the knowledge base and authorized records, then compiles an answer. Every conclusion cites the source document and corresponding section; click to jump to the original text. If the retrieved evidence is insufficient to support an answer, the system clearly states that information is insufficient, never fabricates content.

Simply put: AI only assists with organization and retrieval; conclusions always have a source, and if unsure, it says so directly. For clinical scenarios, this "conservative" approach is actually more practical—not seeking comprehensive answers, but ensuring every statement is reliable and reassuring.

Structured Templates: Standardized Writing, Less Repetitive Work

Repeated writing of similar records, slow onboarding for newcomers, and inconsistent writing standards are common departmental problems.

We organized commonly used structured templates by record type, covering high-frequency documents such as admission notes, first progress notes, daily progress notes, and consultation opinions. Doctors select a template and fill in module by module, instead of building a framework from scratch each time. This saves repeated writing time and ensures structural consistency, making the department's overall record quality more uniform.

Templates are not fixed; they can be adjusted to each department's practice habits. Newcomers get up to speed faster, fill in without missing items, and training costs drop significantly.

Extended Capabilities: Imaging and Lab Visualization

Beyond the documents themselves, we also optimized the browsing experience for test data, so clinicians don't have to open a separate imaging workstation or flip through lab reports.

For DICOM imaging, we support 3D volume rendering and three-plane linked viewing, with axial, coronal, and sagittal views synchronized, no need to jump to another system, making quick reading convenient. All imaging access strictly follows permission validation, only showing data the current user is authorized to see, for auxiliary viewing only, not replacing professional diagnosis.

For lab results, structured data is organized by time and item, displayed with trend charts and scatter plots to show change patterns, with abnormal indicators highlighted. Historical fluctuations and which items are abnormal are much more intuitive than pure numeric reports, easier to compare. Again, all data is based on the original report; charts are for display only, not automatic diagnosis.

Complete Workflow from Writing to Archiving

These capabilities are not piled up; they form a line along the complete lifecycle of a clinical document:

The entire process happens in one workbench, no jumping between multiple systems. Paths are shorter, operations fewer, yet process records are more complete.

Safety and Permissions: The Boundary of Medical Data Must Be Guarded

For medical products, data security and privacy are the bottom line. All features are open only under the premise of permission control and operation traceability.

First, role-based access control. Clinical doctors, department reviewers, QC staff, and administrators have clear boundaries on what they can see, edit, and approve. All permissions are validated at the backend, not just hidden buttons in the frontend; guessing a URL cannot bypass access.

Second, full traceability. Document modifications have version history, workflow operations have process logs, and data access has behavior audits. Who did what at what time, and which patient data was accessed, can all be traced.

Third, restricted retrieval scope. AI Q&A and data queries always include the current user's permission filter, so AI retrieval cannot circumvent business permissions to access unauthorized records.

Fourth, deployment flexibility. Private on-premise deployment is supported; all data stays within the healthcare institution's own environment, not sent to external services, meeting data security and compliance requirements.

Different Roles, Each with Its Own Value

A good system should make every role feel the benefit, not just one end.

Suitable Scenarios

Final Thoughts

We approach products in the medical field with reverence.

We have never believed in "AI replacing doctors," nor do we build flashy features like fully automatic record writing. What we want is a practical tool that reduces burden, improves efficiency, and controls quality—handing repetitive mechanical work to the system and returning professional judgment to doctors.

For medical records, speed is not the goal; stability is. Standardized format, traceable process, AI with evidence, and closed-loop quality control matter far more than fancy concepts. !

Friendly reminder: MedDoc AI output is only for document assistance and data compilation, and does not replace the professional judgment and record review of medical staff. All clinical decisions and formal document confirmations must be completed by qualified professionals.

If you are also working on medical informatization, or have similar product practice and lessons learned in scenarios like record writing, quality control, or collaboration, we welcome discussion and exchange.

Comments

Top 1 of 2 from juejin.cn, machine-translated. The original thread is authoritative.

卷不动了噻 1 likes

Big shot, did you use Java or Node for the backend?

徐小夕

Node.js changes the world.