Value Is in Workflow, Not Replacing Doctors
The most mature healthcare AI does not diagnose autonomously; it removes repetitive work: imaging triage, clinical documentation, follow-up communication, coding and drug candidate screening. These tasks have clear inputs and outputs, human review and interceptable errors. Ask which department and workflow it plugs into, who reviews and who is accountable before comparing model metrics. AI without workflow ownership does not get used, no matter the accuracy.
Imaging and Clinical Support: A Tool Map
Imaging AI focuses on detection and triage for specific conditions: Aidoc prioritizes emergency imaging, InferRead covers pulmonary nodules and stroke, United Imaging AI integrates deeply with scanners, Shukun focuses on cardiovascular and cerebrovascular, Deepwise covers chest and neuro, Airdoc enters chronic disease screening through fundus images and PathAI focuses on pathology. Chinese hospitals should verify NMPA registration, PACS/HIS integration and local support; international organizations also need FDA/CE and reimbursement paths.
Real-time critical-findings AI for radiology
A benchmark in CT lung-nodule AI
The AI suite of an imaging-device giant
Cardio-cerebrovascular imaging AI and a 'digital doctor' portfolio
Multimodal imaging AI covering dozens of diseases in one suite
The first listed AI fundus-screening company
AI pathology for cancer diagnosis
Documentation and Patient Communication: The Underrated Efficiency Layer
Clinical documentation consumes enormous physician time. Nuance DAX, Nabla and Heidi generate draft notes from ambient dictation for physician review and sign-off; MedGPT, HuatuoGPT, Zuoshou Doctor and Huimei fit Chinese records and primary care. For patient communication, AI can handle reminders, post-op follow-up and common questions but cannot replace clinical judgment. Validate speech recognition on dialects and medical terminology, and whether on-premise deployment is supported.
Automatic clinical documentation
Ambient AI scribe that writes clinical notes during consultations
AI medical scribe turning consults into notes, letters and forms
Medlinker's medical LLM for AI diagnosis
Open-source Chinese medical LLM 'HuatuoGPT'
AI triage and consultation assistant for clinics and hospitals
CDSS veteran: real-time AI alerts embedded in doctor workflows
Drug Discovery: Highest Long-Term Value, Hardest to Evaluate Short-Term
AlphaFold transformed protein structure prediction, Isomorphic Labs applies AI to drug design and XtalPi focuses on crystal forms and molecular simulation. Value takes years to validate; do not count generated candidates but wet-lab hit rate, cycle time and cost. Hospitals and pharma companies should define metrics with R&D teams rather than judge by demos. Data quality and the experimental loop matter more than model size.
Procurement and Compliance: Ask About Accountability Before Accuracy
Healthcare AI involves patient safety, privacy and regulation. Before procurement, confirm registration or certification for the intended indication, data location and cross-border transfer, on-premise support, error alerts and human review, and adverse-event traceability. For patient-facing tools, define the boundary of health advice and never treat AI output as diagnosis. The safest pattern is AI triage or draft → clinician review → traceable log, not direct patient-facing conclusions.