In a 2020 study of 100 heart failure patients sent home with a chest patch, a machine learning model flagged trouble a median of 6.5 days before the patient went back into the hospital. That gap is the real role of AI in remote patient monitoring: the devices collect the numbers, while AI decides which numbers matter, which patient needs a call today, and which alert can wait.
Remote patient monitoring (RPM) sends readings such as blood pressure, weight, oxygen levels and heart rhythm from a patient’s home to a care team. AI sits between the device and the nurse. It sorts alerts, spots early decline, nudges patients who stop taking readings and drafts the notes that Medicare billing depends on. Used well, it lets one nurse watch far more patients without missing the ones in danger; used badly, it buries that nurse in noise or hides a sick patient behind a confident score.
LINK-HF Predictions, Fewer False Alerts and 2026 Medicare Codes
- A machine learning model in the LINK-HF study caught 76% to 88% of the warning signs before heart failure readmissions, with 85% specificity.
- Medtronic says its FDA-cleared AccuRhythm AI cut false atrial fibrillation alerts by 74.1% and false pause alerts by 97.4% on the LINQ II heart monitor, while keeping nearly all true alerts.
- Medicare now pays for shorter monitoring months: CPT 99445 covers 2 to 15 days of device data, and 99470 covers the first 10 minutes of management with one live interaction.
- Medicare still requires an FDA-defined medical device, patient consent and an established relationship for RPM. AI review time does not count toward the time-based codes, because they pay for clinical staff and practitioner time.
- The biggest risks are skewed training data, alert fatigue and patient data flowing to vendors, which under HIPAA usually needs a business associate agreement.
How Remote Patient Monitoring Works Before Any AI Is Added
CMS defines remote physiologic monitoring as the use of non-face-to-face technology to track and analyze a patient’s body metrics, such as blood pressure, oxygen saturation, blood sugar and weight. CMS breaks every program into three parts: patient education and device setup, device supply, and treatment management.
The flow is simple. A patient with high blood pressure gets a cellular cuff, and each morning the reading uploads to a secure portal. A nurse reviews the trend, calls the patient when numbers drift, and the doctor adjusts medication. Medicare insists the device collects data electronically and uploads it automatically. A patient typing numbers into an app does not qualify.
The rules around that flow matter for anyone planning an AI layer:
- Established patient: Remote physiologic monitoring requires an existing relationship with the billing practitioner.
- Condition: The program must monitor an acute or chronic condition and be medically necessary.
- Consent: CMS requires patient consent at the time RPM services are provided.
- Device: The device that collects and sends the data must meet the FDA’s definition of a medical device.
- One biller: Only one practitioner can bill remote monitoring for a patient in a 30-day period.
Notice what is missing: nothing in those rules mentions analytics. The AI layer is optional, which is why its value has to show up in staff time, earlier catches and fewer hospital stays.
Where AI Earns Its Place in an RPM Program
A program with 300 patients sending two readings a day produces about 18,000 data points a month. No nurse reads them all. AI helps in four places, and they are worth separating because each has different evidence and different risks.
Alert Triage: Fewer False Alarms, Same True Ones
The oldest RPM software used fixed thresholds. A rule such as “alert when systolic pressure tops 160” fires on a patient who just climbed the stairs and on a patient in real trouble alike. Triage models look at context instead: the patient’s own baseline, the time of day, the last few readings and the signal quality.
The clearest public example comes from cardiac monitoring. In July 2021 Medtronic announced FDA clearance for AccuRhythm AI, two cloud algorithms that check the rhythm events sent by its LINQ II insertable cardiac monitor. Medtronic says the algorithms cut false atrial fibrillation alerts by 74.1% and false pause alerts by 97.4%. They kept 99.3% of true AF alerts and 100% of true pause alerts. Medtronic trained them on more than 1 million heart rhythm episodes from its own database.
Those are company figures from a press release, not a peer-reviewed trial. Still, they show the right way to judge any triage tool. Ask how many false alerts it removes, and then ask how many true alerts it loses. A vendor that quotes only the first number is hiding the one that can hurt a patient.
Deterioration Prediction: Days of Warning, Not Minutes
Prediction goes a step further than triage. Instead of judging one reading, the model learns what normal looks like for each patient and watches for slow drift across many signals at once.
The best-known evidence is the LINK-HF multicenter study, led by Josef Stehlik and published in Circulation: Heart Failure in 2020. Researchers sent 100 heart failure patients home with a disposable chest patch that streamed data through a smartphone to a cloud platform. Machine learning built a personal baseline for each patient. When actual vital signs drifted away from the predicted ones, the system raised an alert.
The model detected the warning signs of heart failure hospitalization with 76% to 88% sensitivity and 85% specificity. The median time between the first alert and readmission was 6.5 days. In practice, a week is enough time to adjust a diuretic dose or bring a patient into clinic. A threshold alarm that fires the night of a crisis cannot offer that.
Adherence: Keeping Patients Taking Readings
RPM only works when patients use the device. Adherence tools use simple models to spot who is drifting away: a patient who skipped two days, a cuff that only syncs on weekends, a scale that stopped reporting after a battery died. The system can send a text reminder, open a support ticket or put the patient on a nurse’s call list.
Adherence now has a direct billing link too. Medicare’s device-supply codes depend on how many days of data a patient sends in 30 days. Practices must collect data for 2 to 15 days or 16 or more days, depending on the code. A model that predicts which patients will fall short helps staff step in before the month closes.
Documentation: Drafting the Notes Behind Every Bill
Time-based RPM codes need a record of who did what and for how long. Generative AI can summarize a month of readings, draft a note after a patient call and pull out the medication change for the chart. The same kinds of tools that write up meetings, covered in our roundup of AI tools for meeting notes, are now built into some care management platforms.
The draft still needs a human reviewer, because a summary model can invent a detail that was never said, which is why some vendors ground their output in the source chart with retrieval. Our explainer on retrieval-augmented generation shows how that grounding works and where it still slips.
Triage, Prediction, Adherence and Notes: What the AI Does
| Use case | What the AI does | Data it needs | Published example | Main risk |
|---|---|---|---|---|
| Alert triage | Scores each alert so true events reach staff first | Device events, patient baseline, signal quality | AccuRhythm AI on LINQ II (FDA cleared, 2021) | Suppressing a true alert |
| Deterioration prediction | Flags slow drift from a patient’s own baseline | Continuous vitals from a wearable | LINK-HF: 6.5-day median lead time | Models trained on narrow groups |
| Adherence support | Predicts who will miss reading days and prompts them | Sync history, reading days per month | Tied to Medicare’s 2 to 15 and 16+ day rules | Over-messaging patients |
| Documentation | Drafts monthly summaries and call notes | Readings, call transcripts, chart | Built into some care platforms | Invented details in notes |
| Workload balancing | Builds a daily call list by risk | Alerts, past calls, staffing | Varies by vendor | Low-risk patients never get a call |
What Medicare Pays for RPM in 2026
Medicare pays for RPM through a set of CPT codes, and the 2026 rules added two new ones. The CY 2026 Physician Fee Schedule final rule adds 99445 and 99470 alongside the older codes in its remote monitoring section. The rule took effect January 1, 2026.
| CPT code | What it covers | Approximate 2026 national rate |
|---|---|---|
| 99453 | Initial setup and patient education on the device | About $22 |
| 99454 | Device supply with daily readings or alerts, 16 to 30 days in 30 days | About $47 |
| 99445 | Device supply with daily readings or alerts, 2 to 15 days in 30 days (new in 2026) | About $47 |
| 99457 | Treatment management, first 20 minutes, with interactive communication | About $52 |
| 99458 | Each additional 20 minutes of treatment management | About $41 |
| 99470 | Treatment management, first 10 minutes, with one real-time interaction in the month (new in 2026) | About $26 |
Rates are approximate and for illustration only. Actual payment varies by location and setting.
The two new codes change the math for AI tools. Before 2026, a patient who sent 12 days of readings earned nothing for device supply. Now 99445 pays for that month. In the same way, a month with only a short check-in call can be billed under 99470. A program that once lost money on patients with patchy use can now cover them, and that makes adherence and triage models easier to justify.
A Worked Example for One Patient Month
Take a patient with high blood pressure who sends readings on 20 days and gets 25 minutes of nurse time, including a phone call. The practice could bill 99454 (about $47) and 99457 (about $52), for roughly $99. If the same patient sends readings on only 10 days and gets a 12-minute call, the practice bills 99445 (about $47) and 99470 (about $26), or roughly $73. Setup under 99453 is billed once at the start, not monthly.
One rule trips up AI vendors’ sales pitches. The time-based codes pay for time spent by clinical staff, physicians or other qualified professionals. Minutes a model spends reading data are not billable minutes. AI saves money by cutting staff time per patient, not by adding billable time.
CMS also lets practices bill RPM alongside chronic care management, principal care management and transitional care management, as long as the same time is not counted twice. A practice cannot bill remote physiologic monitoring and remote therapeutic monitoring for the same patient in the same period.
Risks That Come With AI-Driven Monitoring
Every benefit above has a failure mode. A practice that buys an AI layer should plan for each one before go-live.
Bias in the Training Data
A model is only as broad as the patients it learned from. The LINK-HF cohort had a mean age of 68 and was 98% male. That does not make the results wrong, but it does mean nobody should assume the same accuracy for a 45-year-old woman with heart failure.
Medtronic describes its own training set as “diverse and debiased.” Ask any vendor to show performance broken down by age, sex, race and skin tone, especially for optical sensors such as pulse oximeters. If they cannot, run your own audit on your first few months of data.
Alert Fatigue and Over-Suppression
Alert fatigue is the reason triage AI exists. When staff see hundreds of false alarms, they start clicking through them. The Medtronic figures show how large the false-alert problem can be. Even after filtering, a quarter of the false AF alerts remained.
The opposite failure is quieter and worse. A model tuned to cut noise may also hide true events. Set a written rule for who reviews suppressed alerts and how often, and track missed events as carefully as you track time saved.
HIPAA and Vendor Data Flows
RPM data is protected health information. When a practice shares it with a monitoring platform or an AI vendor that works on its behalf, HIPAA generally treats that vendor as a business associate, which requires a written agreement. Read the contract for how the vendor may use your patients’ data, including whether it can train its models on it.
Staff use matters too. A nurse who pastes a patient’s readings into a public chatbot to “summarize quickly” has just shared PHI with a company that signed nothing. A short internal policy closes that gap. Our template for an AI acceptable use policy is a good starting point for clinics.
Regulatory Gray Areas
Medicare requires the device to meet the FDA’s definition of a medical device. That rule covers the hardware that collects data. It does not mean every analytics layer on top has been cleared. AccuRhythm AI went through FDA clearance. Many dashboard risk scores have not. Ask vendors directly which parts of their product are cleared and for what use.
How to Add AI to an Existing RPM Program
Most practices already run RPM with basic software. Adding AI works best in stages.
- Measure your baseline first: Count alerts per patient per month, the share that led to action, staff minutes per patient and how many patients fall below 16 reading days. Without these, no vendor claim can be checked.
- Start with triage or adherence: Both have clear, countable outcomes. Prediction tools need longer follow-up to prove value.
- Run in shadow mode: Let the model score alerts for four to eight weeks while staff work as usual. Compare its calls with what nurses actually did.
- Ask for subgroup results: Request accuracy by age, sex and race, and the share of true alerts lost at each setting.
- Keep a human on every suppressed alert class: Sample a set number of filtered alerts each week.
- Check the device chain: Confirm every device meets the FDA medical device definition and uploads data automatically, as CMS requires.
Practices choosing hardware can compare makers in our list of top medical equipment manufacturers in the USA. Entrepreneurs looking at the supply side of home care can find related openings in our guide to medical supply business ideas.
Where AI Monitoring Still Falls Short
The evidence base is thinner than the marketing suggests. LINK-HF had 100 patients, and the AccuRhythm results came from the company itself. Both are useful, and neither proves that AI-driven RPM reduces deaths or total costs across a health system.
Our view is that the best case today is narrow and practical: fewer false alarms, earlier calls for patients who are drifting, and less time on paperwork. Those gains are measurable in a single clinic within a quarter. Broad claims about AI replacing clinical judgment in home monitoring are not supported by the published data, and practices should treat any vendor that makes them with care.
Billing AI Time, FDA Clearance and HIPAA for Monitoring Data
What Is AI in Remote Patient Monitoring?
It is software that analyzes data from home medical devices, such as blood pressure cuffs, scales and heart monitors. It sorts alerts by urgency, predicts which patients may get worse, prompts patients who miss readings and drafts clinical notes. Nurses and doctors still make the care decisions.
What Are the CPT Codes for Remote Patient Monitoring in 2026?
The main codes are 99453 for setup, 99454 for 16 to 30 days of device data, 99445 for 2 to 15 days, 99457 for the first 20 minutes of management, 99458 for each extra 20 minutes and 99470 for the first 10 minutes. 99445 and 99470 are new for 2026.
Can AI Time Be Billed Under RPM Codes?
No. Codes 99457, 99458 and 99470 pay for time spent by clinical staff, physicians or other qualified health professionals. Time a model spends reviewing data does not count. AI lowers the staff time needed per patient instead.
Are AI Remote Monitoring Tools FDA Cleared?
Some are. Medtronic’s AccuRhythm AI algorithms for the LINQ II cardiac monitor received FDA clearance in 2021. Many risk scores inside RPM dashboards have not been cleared, so ask each vendor which features have FDA clearance and for what use.
Does HIPAA Apply to Remote Patient Monitoring Data?
Yes. RPM readings are protected health information when a covered provider collects them. Vendors that handle the data for the provider generally need a business associate agreement, and staff should not paste patient data into public AI tools.
Can AI Predict Heart Failure Before a Hospital Visit?
In the LINK-HF study, a machine learning model using a wearable chest patch detected warning signs with 76% to 88% sensitivity. The first alert came a median of 6.5 days before readmission. The study was small, with 100 mostly male patients, so results may differ in other groups.
CMS Fee Schedule Rules, the LINK-HF Study and Medtronic’s Data
- CMS: Telehealth and Remote Patient Monitoring, MLN901705 (December 2025)
- Federal Register: CY 2026 Physician Fee Schedule Final Rule (November 5, 2025)
- CMS: CY 2026 Physician Fee Schedule Final Rule Fact Sheet
- Prevounce: 2026 Remote Patient Monitoring CPT Codes
- Stehlik et al.: Continuous Wearable Monitoring Analytics Predict Heart Failure Hospitalization, the LINK-HF Multicenter Study (Circulation: Heart Failure, 2020)
- Medtronic: FDA Clearance and Results of AccuRhythm AI Algorithms (July 28, 2021)
- HHS: Business Associates Under HIPAA
