Day 2 · R2 Smart Healthcare
Three speakers share the hard-won lessons of deploying smart biomedical platforms in clinical settings and government AI-adoption policy, covering agentic LLM clinical applications, decades of hospital information-system evolution, and practical AI adoption on the nursing floor.
At a glance
| Speaker | Talk | In one sentence |
|---|---|---|
| Kao Hung-Yu (moderator) | Opening | Three speakers share the hard-won lessons of deploying smart biomedical platforms in clinical settings, along with government AI adoption experience. |
| Hsu Kai-cheng (許凱程), Chief Medical Officer, Industrial Technology Research Institute (ITRI) | Agentic LLM and Smart Healthcare | Uses agentic LLMs to chain data integration, SaMD (Software as Medical Device), and clinical deployment into automated task chains, targeting clinical settings where it is "hard even to get a physician to move one extra finger." |
| Lai Lai-hsun (賴來勳), Director, Information Management Office, Taichung Veterans General Hospital | From Digital Transformation to Smart Hospital | 40 years of hospital information system evolution; ranked Taiwan's No. 1 in Newsweek's World's Best Smart Hospitals for three consecutive years, 2023–2025. |
| Kang Shih-Chung (康仕仲), CEO, Jubo Health (智齡科技) | How can AI help nurses? | Nurses spend 60% of their time writing records instead of giving care; Jubo uses mobile devices plus lightweight AI and LLM+RAG (retrieval-augmented generation) to boost care operation efficiency. |
3 of 3 talks
1Agentic LLM and Smart HealthcareHsu Kai-cheng (許凱程), Chief Medical Officer, Industrial Technology Research Institute (ITRI)
Uses agentic LLMs to link biomedical data integration, SaMD (Software as Medical Device), and clinical task decomposition together, enabling AI to truly take root in hospitals.
Key points
- Opened by contrasting agentic LLMs with traditional LLMs: a traditional LLM (like ChatGPT) handles a single command for a single task, while an agentic LLM is like Iron Man's butler Jarvis — give it a goal (e.g. "plan a trip to Japan") and it automatically decomposes it into sub-tasks such as booking flights, lodging, and transportation, finds tools to complete each one, then integrates the results; the human only needs to confirm whether the result is correct, since "it's impossible to let AI make the decisions itself."
- The biggest obstacle to medical data integration is privacy: a Constitutional Court (Grand Justices) interpretation requires a dedicated special law before medical data can be released for research use; the speaker said the special law is expected to be enacted by August 2025, and until then data flow remains tightly restricted.
- SaMD (Software as Medical Device) must obtain regulatory clearance from the TFDA's Smart Medical Device Office according to device classification (Class I/II/III), including categories such as CADx (computer-aided diagnosis) and CAD-triage (computer-aided triage); this is described as the "foundation" of smart healthcare.
- Presented the Ministry of Health and Welfare's "three arrows" policy: (1) an AI Trustworthy Center (opening up the AI black box), (2) a regulatory clearance center (helping hospitals get AI models cleared), and (3) an NHI (National Health Insurance) sandbox (evaluating medical AI's benefit before and after go-live — e.g. mortality, processing speed, changes in clinical staff loading — with the prospect of NHI reimbursement if the benefit is large enough) — bluntly noting that many hospitals' current AI applications are effectively "volunteer work," with no one paying for them.
- Nursing staff shortages were a pain point running through the whole talk: some hospitals have 80% of wards "able to open but closed" due to insufficient nursing staff; the purpose of smart healthcare tools is not to replace staff but to ease their burden, keeping more nurses from moving to lower-workload jobs in the cosmetic medicine industry.
- Showcased already-deployed cases:
- Non-contrast CT that directly predicts the extent of ischemic stroke lesions: for lesions >70mL, accuracy is close to 80%, and detecting the presence of infarction can reach 90% — in some regions even outperforming commercial software (more accurate when compared against MRI ground truth).
- Ambulance AIoT: installs an AI diagnostic module on ambulances, shortening pre-hospital delay from about 2 hours to just over ten minutes.
- Home single-lead ECG sleep monitoring AIoT: replaces the traditional polysomnography (PSG), which requires waiting over a month and an overnight stay in an unfamiliar setting — it can read sleep staging (including REM), apnea, and arousal events from a single-lead ECG, and has already won a National Innovation Award.
- Multimodal stroke-risk AI: integrates physiological signals such as blood pressure, heart rate, and respiration to predict complications like recurrent stroke, aspiration pneumonia from dysphagia, and embolism from atrial fibrillation.
- The "Chiayi Big Platform," a collaboration with the National Clinics Association: collects data via wearables and home glucose meters, imports it into the EMR, then uses a dedicated LLM + RAG to assess risks such as diabetic nephropathy and retinopathy and give treatment/dietary recommendations — aimed at closing the gap between clinics' far weaker data-collection capacity and that of hospitals.
- Hospital at Home / HomeSpital: paired with the Ministry of Health and Welfare's home-based acute care policy (for pneumonia, cellulitis, etc.) that took effect in July 2024; since caregivers are not medical professionals, AIoT assistance is even more needed to judge whether a patient must be transferred to hospital.
- Decentralized clinical trials (DCT): uses AIoT to let trial participants transmit physiological data from home, reducing the frequency of return visits and improving compliance and trial success rates; both the US FDA and Taiwan's TFDA have voiced support for this direction.
- The BOBI system (a voice-based medical documentation LLM tool): shortened initial-visit documentation time from 20–30 minutes to 5–10 minutes, and shortened nursing shift-handover documentation from 1 hour after the shift ends to about 15 minutes. To address the speech-recognition pain point of Chinese-English code-switching and heavy use of medical abbreviations in clinical settings, after training on specialized corpora the error rate dropped from 30–40% to below 6%, the threshold needed to be clinically usable.
- Presented a real-world test case of agentic task decomposition (in collaboration with National Taiwan University): for an X-ray of an ER fall patient from September 2020, feeding the image directly to a multimodal model with no task decomposition produced an error rate as high as 50–60%; switching to a decomposed workflow — first judge the clinical situation, then judge the image type, then judge the findings within the image — raised accuracy to over 90%.
- In the data-processing pipeline, standardized codes such as ICD-9, SNOMED, and LOINC are used first to organize the clinical database (because clinical records often contain biased noise), and only then is it handed to the language model for final synthesis — ensuring that only "the correct database plus the correct task input" can produce a correct result.
- LLM + RAG has already been applied to a dementia care Q&A platform (answers are annotated with sources from specialized databases to reduce hallucination), GPT-4o multimodal X-ray fracture reading plus treatment recommendations, and vascular physiology interpretation.
Tech, products & figures
- Agentic LLM
- An LLM application pattern with the ability to autonomously decompose tasks, select tools, and integrate results, as distinct from single-turn, command-style LLMs.
- SaMD (Software as Medical Device)
- TFDA's smart medical device clearance classification, with CADx (computer-aided diagnosis) and CAD-triage (computer-aided triage) as representative sub-categories.
- Ministry of Health and Welfare's "Three Arrows"
- AI Trustworthy Center, a regulatory clearance center, and an NHI (National Health Insurance) sandbox.
- BOBI
- A voice-based medical documentation LLM tool developed internally by ITRI / Gongyi Yuan (公益院), for outpatient visits, nursing shift handovers, and meeting minutes.
- Non-contrast CT stroke lesion prediction model
- Estimates the extent of a stroke lesion without needing a contrast agent injection; accuracy ranges from 70% to over 90% depending on lesion size.
- Single-lead ECG sleep analysis AIoT
- Replaces the traditional overnight polysomnography (PSG).
Notable quotes
If you ask him to open a separate window, go somewhere else and open up another computer, that's absolutely impossible — the chance of it actually being adopted... the success rate is extremely low.
4.5 hours, 8 hours — that sounds like quite a while... but it's actually not, because by the time the patient is brought to the hospital, 3 hours have already passed, so you may only have 1.5 hours left to make a fast decision.
Q&A
- No Q&A (the moderator went straight into introducing the next speaker).
Fact-check notes
- Hsu Kai-cheng moved to become President and Chief Medical Officer of Acer Healthcare (宏碁智醫) in April 2025; at the time of this talk (September 2024) he was employed at ITRI, consistent with the official agenda listing. His background (National Taiwan University medical degree, a master's in electrical engineering and a PhD in chemical engineering from National Chung Cheng University, former Director of the AI Center at China Medical University Hospital, and former NIH researcher in the US) is consistent with his own description in the talk ("previously the Director of the AI Center at China Medical University Hospital").Sources：經濟日報台灣醫療科技展新聞
- VeriOsteo OP (智骨篩): an AI developed jointly by Taichung Veterans General Hospital and Acer Healthcare for opportunistic osteoporosis screening from a single chest X-ray; it has obtained TFDA clearance. Publicly reported clinical validation figures are approximately 88.7% sensitivity and 89.4% specificity (which differs somewhat from the "98% accuracy" figure the speaker mentions in the next talk — possibly due to a different validation set, or a slip of the tongue).Sources：GeneOnlineCIO Taiwan
2From Digital Transformation to Smart HospitalLai Lai-hsun (賴來勳), Director, Information Management Office, Taichung Veterans General Hospital
Traces Taichung Veterans General Hospital's 40 years of information-system evolution; ranked Taiwan's No. 1 in Newsweek's World's Best Smart Hospitals for three consecutive years, 2023–2025.
Key points
- On site the hospital is commonly called by its short name, "台中榮總" (Taichung Veterans General Hospital); its official full name is "臺中榮民總醫院" (Taichung Veterans General Hospital, written out in full).
- Hospital scale: founded in 1982, with over 40 years of history; the only national-level public medical center in central Taiwan. Its system comprises 1 medical center plus 3 regional/district hospitals plus 4 veterans health service groups, forming a three-tier medical system from medical center down to clinics.
- 2023 service volume (a post-pandemic surge year): 1,632 beds, nearly 6,000 staff, 124 outpatient specialties, and over 3 million patients served; more than 9,300 outpatient visits per day, 200 ER visits per day, 6,500 inpatient admissions per month, and over 4,800 surgeries per month — all up more than 20% from 2019 — while the number of beds and staff did not increase, posing a huge challenge for the information system.
- Awards and international rankings: ranked Taiwan's No. 1 for three consecutive years (the 2023–2025 lists) in Newsweek's World's Best Smart Hospitals ranking; on the list for the 2025 edition (published around the time of this talk), it broke into the global top 100 for the first time, at No. 99 — the speaker explained that ranks 101–350 are not individually ranked and are instead listed alphabetically, so No. 99 is one of the last numerically-ranked positions before that unranked tail. Separately, in an early-2024 HIMSS DHI (Digital Health Indicator) assessment it ranked 2nd in the world, with 1st place going to China Medical University Hospital, where Hsu Kai-cheng previously worked.
- In 2023, passed both HIMSS EMRAM Stage 6 and Stage 7 assessments within 8 months, reportedly the fastest time any hospital in Taiwan's medical community has passed both stages.
- History of information-system development: 1982, IBM mainframe systems (outpatient/inpatient, the PEX system) → 1999–2010, moved off the IBM mainframe to open systems, rebuilt the outpatient/inpatient system and added features such as physician summaries and nursing actions → 2011, in step with Department of Health (today's Ministry of Health and Welfare) policy, implemented electronic signatures and medical-record digitization, and developed connections to medical instruments (now called medical IoT, linking anesthesia, respiratory, and vital-sign measurement devices) → from 2017, integrated hospital-wide data to build a data warehouse and dashboards → after Superintendent Chen (Chun-chieh) took office in 2022, put full effort into developing AI models.
- A patient-centered inpatient system: a single screen integrates the patient's basic information, current medications, ongoing treatments, allergy history, latest nursing records, shift handover notes, and latest vital signs, with one-click links to detail screens of other subsystems such as ventilators — the speaker emphasized this design was refined through repeated iteration with clinical physicians.
- Structured medical records were pushed in two phases: as early as 2004, admission/discharge summaries, progress notes, and nursing records were already structured (checkbox fields plus retained free text); after Superintendent Chen took office he further questioned "why is the report still a piece of prose," prompting a project that spent a year meeting department by department to consolidate report formats originally tied to over 3,000 billing codes into 615 standardized report templates (covering endoscopy, radiology, and pathology reports), letting physicians produce reports via checkboxes while still keeping the flexibility of free text.
- Electronic indicator label medication boxes: replace handwritten/sticker medication box labels; after the improvement, medication labels for an entire ward (illustrated with a 70-bed ward) can be updated in about 3–5 minutes, with dual verification by scanning both the medication and the patient's QR code before dispensing — solving the old problem that medication labels could not be computer-audited and were prone to falling off or becoming illegible.
- Full-color electronic bedside cards: upgraded from three-color (black/white/red) e-paper across more than 400 beds (starting around 2018–2019) to fully full-color this year (2024), with image update time shortened from about 30 seconds to about 1 second; content updates automatically in real time as physician orders and nursing records are entered.
- Taichung Veterans General Hospital was the first hospital in Taiwan to legally and fully implement electronic medical records (electronic nursing records), around ROC year 98, i.e. circa 2009.
- Telemedicine: video outpatient consultations grew substantially during the pandemic, but the original rules lapsed once the pandemic ended; a new telemedicine regulation (covering a different scope of eligible patients than pandemic-era video consultations) took effect on 1 July 2024. Superintendent Chen invested substantial resources to build what is billed as "Taiwan's largest telemedicine center" (established in 2022); even though the NHI does not currently reimburse it and it is not profitable to run, it is still seen as an inevitable trend, since hospital beds and staffing are limited while demand for elderly care keeps rising, requiring a Hospital at Home solution. It has already connected more than 25 sites (partner hospitals, branch hospitals, veterans health service groups, and rural health stations).
- As early as 2021, partnered with Chunghwa Telecom to combine a dedicated 5G network with AR glasses for broadcasting cardiac catheterization procedures and integrating with da Vinci surgical systems, testing the feasibility of 5G medical applications; then-President Tsai Ing-wen held a video conference from the medical technology exhibition floor with a Taichung Veterans General Hospital surgical team 200 km away; the Discovery Channel also ran a feature on Taiwan's integration of IT with medicine.
- A full-journey monitoring system for referred patients: for patients being transferred from a branch hospital to the main Taichung campus (cardiac patients, stroke patients, unstable newborns, etc.), it monitors vital signs throughout the transfer, live-records inside the vehicle, and uses Google Maps to track the estimated arrival time so the ER team can stand by in advance; more than 500 referral cases have been completed to date. Because the main Taichung campus was the earliest (2012) to complete its core information-system transformation, the Veterans Affairs Council chose this system as a model, and it has now been rolled out to all branch hospitals in Taiwan's veterans-hospital system (including the Taipei and Kaohsiung Veterans General Hospital branches).
- During the 2020 pandemic, completed system rollout to 8 branch hospitals within a single month: each rollout mobilized at least 40 people on-site (IT, biomedical engineering, nursing, laboratory), with no one infected during the process and no branch hospital's rollout failing or being rolled back — all succeeded on the first attempt — which the speaker credited to strict discipline and management.
- Won the Joint Commission of Taiwan's institution-wide Smart Hospital award three times: in 2015, 2019, and 2023.
- After taking office, Superintendent Chen reorganized what had been the "Big Data Management Committee" into the "Smart Healthcare Committee," divided into about 8 working groups running multiple internal projects in parallel, alongside partnerships with universities in central Taiwan and nationwide, plus some industry collaborations that cannot be disclosed.
- Results from a national program: undertook the National Science and Technology Council's "Taiwan Smart Healthcare Alliance" (TSHA) program, a three-year effort from 2021–2024 that has now concluded, with Taichung Veterans General Hospital serving as the overall coordinator, integrating AI models from what the speaker described as about 19 medical centers nationwide and advancing federated learning and cross-hospital validation. Two of its own AI models: an osteoporosis-screening AI called "VeriOsteo" developed with Acer (judges osteoporosis from a single X-ray, with the speaker citing about 98% accuracy), and a pressure-injury (bedsore) assessment-assist model developed with Tunghai University and Hui De Technology — both are said to have obtained FDA clearance.
- Currently hosts two of the three national-level centers under the Ministry of Health and Welfare's "three arrows" policy: the "Clinical AI Regulatory Validation Center" that interfaces with TFDA (including 4 cross-hospital validation cases and 3 final clearance cases), and the "AI Impact Research Center" that interfaces with the National Health Insurance Administration (assessing how much NHI reimbursement AI adoption should receive).
- AI application examples: Healthcare GPT (a collaboration with National Yang Ming Chiao Tung University this year) automatically generates admission medical record summaries (AutoMission Note); ChatGPT-like tools help draft reply letters to complaints submitted through the superintendent's mailbox (manually drafted replies were previously often rejected/returned); an X-ray positioning verification AI detects when the actual positioning doesn't match the physician's order and immediately flags for a retake, preventing patients from receiving extra radiation exposure due to positioning errors.
- International exchange: holds an international smart healthcare training program every year, with long-term cooperation agreements already signed, drawing visitors from about 40 countries annually; on 9 August 2024, President Lai Ching-te visited the hospital to see its smart healthcare achievements and expressed surprise and delight at how much progress had been made.
Tech, products & figures
- HIMSS EMRAM / DHI
- Electronic Medical Record Adoption Model (EMRAM) and Digital Health Indicator (DHI) assessments from the US-based Healthcare Information and Management Systems Society (HIMSS).
- Newsweek World's Best Smart Hospitals
- Newsweek's World's Best Smart Hospitals ranking.
- VeriOsteo (智骨篩)
- An X-ray osteoporosis-screening AI developed with Acer Healthcare (see the verification note in the previous talk).
- Electronic indicator label medication boxes / full-color electronic bedside cards
- E-paper systems for inpatient medication and bedside information, both of which have patents pending; the electronic indicator label medication box method was later replicated free of charge by the National Health Insurance Administration and rolled out for use by clinic physicians nationwide.
- Taiwan Smart Healthcare Alliance (TSHA)
- A three-year (2021–2024) cross-hospital AI validation and clearance program run by the National Science and Technology Council.
Notable quotes
There's no money in doing this, because the NHI doesn't reimburse it right now... but it's a trend, so we just have to start first.
Not a single hospital's rollout failed or was rolled back — every one succeeded on the first try.
Q&A
- No Q&A.
Fact-check notes
- Newsweek ranking: public reporting shows Taichung Veterans General Hospital ranked 246th globally on the 2024 list (the only hospital from Taiwan, China, or Hong Kong in the top 250), broke into the global top 100 for the first time on the corresponding 2025-edition list, and improved to 85th on the 2026-edition list — consistent in timing with the speaker's statements that it was "Taiwan's No. 1 for three straight years, 2023, 2024, and 2025" and that "this year it broke into the top 100 at No. 99" (Newsweek's list is conventionally named for the following year and published each autumn).Sources：GeneOnline健康醫療網
- HIMSS DHI ranking of 2nd worldwide, with China Medical University Hospital 1st: consistent with public reporting.Sources：中國醫藥大學附設醫院官網臺中榮總官網
- Taiwan Smart Healthcare Alliance (TSHA): verification found the official program described as a national AI team formed from "23 cross-hospital teams made up of medical centers nationwide," which does not fully match the speaker's stated figure of "19 medical centers" — this may be a discrepancy between the number of teams versus the number of participating medical centers, and it could not be precisely reconciled; the program's timeline (proposed in 2021, concluded and presented in 2024) matches what the speaker described.Sources：聯合新聞網台北醫學大學計畫頁
- EMRAM Stage 7: verification found that Taichung Veterans General Hospital was the first public hospital to pass the HIMSS EMRAM Stage 7 international assessment, but no independent third-party reporting could be found to corroborate the exact timeline of "passing Stage 6 and Stage 7 within 8 months"; this is listed provisionally (not fully verified, speaker's own account only).
3How can AI help nurses?Kang Shih-Chung (康仕仲), CEO, Jubo Health (智齡科技)
Nurses and care attendants spend 60% of their working hours writing records; Jubo uses mobile devices plus lightweight AI and LLM+RAG (retrieval-augmented generation) to boost operational efficiency in long-term care.
Key points
- Speaker's background: holds a PhD from Stanford University, was formerly a professor in National Taiwan University's Department of Civil Engineering and Deputy Dean of Academic Affairs, was later recruited by the University of Alberta in Canada (once cited as an emblematic case of Taiwan's academic brain drain), and left academia in 2018 to found Jubo Health (智齡科技). The speaker also mentioned being one of about 10 founding members from the founding era of the "Data Science Conference," the predecessor of this conference.
- Core motivating data: Taiwan's old-age dependency ratio (population aged 65+ versus the 15–65 working-age dependent population) currently stands at about 1:4; the speaker cited projections of about 1:2 around 2044 and 1:1 by 2060, and used this data more than a decade ago to foresee that nursing/care-worker shortages would keep worsening. This is now happening simultaneously in Taiwan, Canada, the US, and Japan — hospitals are closing beds, and outpatient services are also starting to shrink.
- A time study conducted while teaching at NTU (in collaboration with faculty from the design department and medical school, embedded to observe care sites in person): care attendants spent only about 40% of their time on direct care, with 60% spent writing records, making phone calls, and doing administrative work.
- Three stages of nursing records: (1) nursing assessment (periodic evaluation of cognition, chewing/swallowing, nutrition, mobility, etc.), (2) nursing plan (care planning), and (3) execution records (blood pressure/glucose measurements, diet, toileting/output volume, mood and activity, etc.) — much of which is unstructured data.
- The founding thesis: there is a need for an AI-powered efficiency support system for elderly care, targeting nursing staff at long-term care institutions, day-care centers, and home care providers outside the hospital system, whose information-system resources lag far behind hospitals — their system complexity is no less than a hospital's, yet their staffing and resources are extremely limited.
- Market status: the speaker's field surveys found that about one-third of hospitals/long-term care institutions across Taiwan still use handwritten nursing records (hand-drawn TPR charts), and the Japanese market has an even larger figure — about 2,500 hospitals still using paper nursing records.
- Product positioning is not a simple data repository but a "care collaboration and communication platform" (comparable to Slack or Teams), designed mobile-first (phones/iPads rather than the nursing station PC), and compliant with HIPAA-level information-security requirements for the US and Canadian markets (matching the mobile-security standards used by banking apps).
- Three main product lines:
- Mobile nursing information system: lets nurses record directly at the bedside using a phone/iPad, replacing having to queue for a computer at the nursing station; because it already complies with mobile-device security standards, it naturally resolves in-hospital vs. out-of-hospital network and security concerns in "hospital at home" scenarios.
- A small vital-signs measurement cart: connects via Bluetooth to many different brands of measurement devices on the market, with results automatically uploaded to the cloud and checked by AI in real time; already deployed in the US, Canadian, and Japanese markets (in Japan it has even been adapted for use in dental clinics).
- FamilyLine (family contact platform): replaces the process where institution staff manually "translate" professional nursing records into language families can understand and then communicate it one by one via LINE; its rationale draws on a paper by John Huskin — that medical disputes mostly stem from insufficient mutual trust between patients/families and providers, and the solution is "more communication," but without adding to nurses' workload.
- Key points from a customer testimonial video (interviews with long-term care/day-care operators): before adoption, nurses/care attendants often resigned due to being overloaded with paperwork (especially handwritten contact notebooks); shift handovers and service records no longer require gathering for 30–60 minutes daily to process; attractive statistical reports can be produced quickly for accreditation reviews; the system can be trained/taught separately by role — social worker, nurse, care attendant — leading to higher satisfaction; even migrant caregivers are able to operate it.
- Company scale: the largest market share among long-term care information platforms in Taiwan, with over 1,200 institutional clients (including overseas), and personalized vital-sign models serving about 130,000 elderly individuals in Taiwan.
- The speaker divides AI's contribution to healthcare into three categories: diagnosis, treatment, and operation (operational efficiency) — and said that as a data scientist, the area he finds most interesting and is most invested in is operation.
- Four AI features already live:
- Vital-sign anomaly alert AI: using very simple unsupervised learning, builds a small personalized vital-sign model for each of the 130,000 elderly individuals; measurements only get a binary "abnormal or not" judgment (no diagnosis), and an abnormal reading is flagged with an exclamation mark for a nurse to review; after going live, the vital-sign recording rate rose from a theoretical 58% by nearly 40 percentage points, and about 72% of abnormal alerts prompt the nurse to open a follow-up nursing record/action; it has been live for over 3 years.
- Wound-care AI: switches to tablet photography, with AI estimating wound size and depth, replacing the old workflow of measuring with a ruler, probing depth with a cotton swab, attaching labels, taking photos, and manually transcribing data exported from an SD card — a process that took 15–20 minutes; the new process takes about 5 minutes and is touchless. It uses a human-in-the-loop approach to keep refining the model; agreement between the AI's and nurses' severity judgments is about 73.5% (close to the roughly 70–80% agreement seen between different human raters, since severity assessment inherently has a subjective component); it has been live for nearly 2 years.
- AI-generated FamilyLine family contact notes ("An Xin Bao Family Contact Notebook," not yet officially live at the time of this talk): a survey found that day-care center staff spend an average of 71 minutes a day writing family contact notes (needing to translate professional records — blood pressure, diet, activity, mood, hygiene, etc. — into language families can understand, communicate it via LINE, and keep tracking family follow-up questions); institutions without this system often resort to using personal LINE accounts. The new feature uses generative AI to automatically produce family-readable text from existing care records in about 3 minutes; staff can fine-tune it and pick a preferred version, and the system records which prompts/outputs get adopted more often to keep improving model selection. Data-governance design: to avoid sending elders' raw personal data directly to an external public LLM, the approach first performs data governance (de-identification, extracting only necessary fields), calls an external/public LLM to process the de-identified content, then reattaches the private fields afterward, and deliberately makes the LLM "a bit dumber" to reduce hallucination and avoid it confidently giving wrong answers. Cost consideration: LLM API fees compress commercial margins, so the team continuously tests multiple LLMs (including cheaper options or self-built in-house small models) to balance cost and quality.
- N-Copilot (安口拍, launching in 2025): inspired by the "Nursing Stanford" concept, it targets the pain point where long-term care nursing staff are frequently interrupted by families (e.g. "When can my mom see Dr. Zhang again?"), doctors (e.g. "Has the grandma in bed 12 had her blood sugar checked, does she need insulin?"), or residents with impromptu questions, and must dig back through days-to-weeks of chart data on the spot to answer — interrupting work, lowering efficiency, hurting the experience, and making hasty, error-prone answers more likely. N-Copilot is a Copilot/ChatGPT-like chat interface built on RAG over the institution's own nursing records; its answers come with traceable source data, and it is designed so that when the database doesn't support an answer it "cannot just give an answer casually the way a person might" — i.e., it must not fabricate. The speaker said the company's internal AI team has effectively morphed into a "data governance team" rather than simply an "AI team," because the real difficulty is data governance, fast and precise retrieval, and privacy protection before ever connecting to the LLM — not the model itself.
- Closing thoughts: introducing AI into nursing operations saves time for government accreditation reviews, for families, for the care recipients themselves, and for institutions' operational statistics management and handover meetings; the company serves more than 1,000 long-term care institutions; the speaker believes generative-AI autonomous agents can not only reduce the existing staff workload but also help new nurses onboard quickly and learn faster how to do care work correctly — which is especially important given the ongoing loss of nursing staff.
Tech, products & figures
- Jubo (智齡科技)
- A long-term care information platform with the largest market share in Taiwan, serving 1,200+ institutional clients and running personalized vital-sign models for 130,000 elderly individuals.
- Vital Sign AI Alert
- Uses unsupervised learning to build a small, personalized vital-sign anomaly-detection model for each resident.
- Wound-care AI (WoundCare AI)
- Estimates wound size/depth from tablet photographs.
- FamilyLine (family contact department) / An Xin Bao Family Contact Notebook
- AI automatically generates family-readable care summaries and sends them via LINE.
- N-Copilot (安口拍)
- A RAG-based Q&A assistant built on an institution's own nursing records, launching in 2025, inspired by the "Nursing Stanford" concept.
Notable quotes
A nurse isn't there to process data — when we train a good nurse, they should be at the bedside doing a lot of bedside care; the nursing data that comes after that really needs good tools to help them with it.
We gradually came to feel that we'd been reshaping our AI team into... by the end it's really more like a Data Governance team, not a team that does AI.
Q&A
- No Q&A (the speaker moved on immediately after closing thanks).
Fact-check notes
- Kang Shih-Chung's academic and career background (Stanford PhD, former NTU Civil Engineering professor and Deputy Dean of Academic Affairs, moved to the University of Alberta in Canada, founded Jubo Health/智齡科技 in 2018) is consistent with public reporting.Sources：博客來書籍介紹天下雜誌Jubo 官網
- The company-scale figures cited ("1,200+ institutions, 130,000 elderly individuals") are broadly consistent in magnitude with public reporting from around 2024 of "over 1,200 long-term care institutions, health records for over 150,000 residents"; the speaker's stated figures are somewhat conservative, and no clear contradiction was found.Sources：Taiwan Tech Arena
- "N-Copilot / 安口拍" and the "FamilyLine family contact notebook AI" are product names the speaker announced during the talk for not-yet-live or just-launched features; no independent third-party reporting could be found in public sources corroborating their functional details or launch timeline, so this is marked as unverified, speaker's own account only.