Industries / Health / Healthcare
Emergency care
Treating sudden illness and injury, from the first minutes at the scene and in the ambulance to the hospital's emergency department.
7
Milestones achieved, from 1792 to 2023.
3
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1792
Ambulances and triage
Dominique-Jean Larrey, a surgeon in the French army, set up two- and four-wheeled wagons pulled by teams of horses to carry the wounded quickly from the battlefield. He argued that the wounded should be treated in order of how badly they were hurt, without regard to rank, a sorting now called triage.
Credit
Dominique-Jean Larrey
1960
Pressing on the chest
W. B. Kouwenhoven, J. R. Jude and G. G. Knickerbocker published "closed-chest cardiac massage" in JAMA in 1960: pushing on the chest by hand to keep blood moving when the heart has stopped. Knickerbocker had noticed, while testing a defibrillator on dogs, that the weight of its paddles on the chest raised blood pressure. The team said the method could restore about half of normal blood flow, and by the end of the 1960s it had replaced older first-aid methods under the name cardiopulmonary resuscitation (CPR).
Credit
William B. Kouwenhoven, James R. Jude, Guy G. Knickerbocker
1967
Heart care in the ambulance
In 1965 Frank Pantridge, a physician in Belfast, built a defibrillator, a machine that shocks the heart back into a normal rhythm, that ran on car batteries and weighed 70 kg, and put it in an ambulance. His BMJ obituary says this created pre-hospital coronary care. With John Geddes he reported the mobile unit in The Lancet in 1967. The approach was taken up quickly in the United States, and Pantridge later designed a version weighing 3 kg.
Credit
J. F. Pantridge, J. S. Geddes
1978
One order for injuries
Advanced Trauma Life Support (ATLS), a course first taught in 1978 and now run by the American College of Surgeons, teaches a fixed order for assessing and treating an injured person in the first hour. Its creator, the orthopedic surgeon James Styner, built it after his family received poor first care following a plane crash in 1976. The college says the course is taught in more than 80 countries and has trained more than 1 million clinicians.
Credit
James Styner; American College of Surgeons
2004
Defibrillators for the public
In a trial across 24 regions of North America, more than 19,000 volunteers in 993 community units were trained either in CPR alone or in CPR plus an automated external defibrillator (AED), a portable box that reads the heart's rhythm and tells the user when to shock. Units with AEDs had twice as many patients survive to leave hospital: 30 of 128 cardiac arrests, against 15 of 107.
Credit
A. P. Hallstrom and the Public Access Defibrillation Trial Investigators
2021
An alert that changed little
In Copenhagen a machine-learning model listened to 169,049 emergency calls from September 2018 to December 2019 and flagged 5,242 as possible cardiac arrests. For half of the flagged calls, chosen at random, the dispatcher saw an alert. On its own the model caught more arrests than dispatchers did (85.0% against 77.5%), with many more false alarms, but dispatchers who saw the alert did not recognize significantly more arrests (93.1% against 90.5%).
LessonA model that beats people in a test does not help unless it changes what people do.
Credit
S. N. Blomberg, F. Lippert and colleagues, Copenhagen Emergency Medical Services
2023
Defibrillators by drone
In Sweden, drones carrying AEDs were sent to suspected cardiac arrests. Of 211 alerts, a drone flew in 72 and delivered an AED in 58. In 37 of the 55 cases with both arrival times recorded (67%), the drone arrived before the ambulance, by a median of 3 minutes 14 seconds. Six drone-delivered AEDs were attached, and 2 patients were shocked before the ambulance came.
Credit
S. Schierbeck, A. Claesson and colleagues, in The Lancet Digital Health
Watch for
An AI coach for bystanders
ChatCPR is a free, open-source AI agent, built on 911 dispatcher training materials, that coaches a caller through CPR. On 12 real, publicly available 911 call recordings, its instructions covered 100% of basic CPR steps and 99% of advanced steps, against 85% and 63% for the dispatchers on those calls. General AI models tested on the same checklist averaged 90% on basic steps and 70% on advanced steps.
Achieved whenan outside group shows, on manikins or on real calls, that people coached by it give better CPR than people coached by dispatchers.
Claimed by
N. Desai, J. W. Ayers and colleagues at UC San Diego, in JAMA Internal Medicine, May 2026
AI reads heart-attack ECGs
The FDA granted De Novo authorization, its route for a new type of device, to Powerful Medical's Queen of Hearts. It is an AI model that reads an ECG and flags heart attacks that need a cardiologist at once (STEMI and STEMI equivalents). In a retrospective registry study across three US heart-attack networks, false alarms that called in the catheter-lab team were 7.9% with the tool against 41.8% without it; the study's lead investigator reports research support from the company.
Achieved whena study without company funding shows, in routine emergency use, fewer missed heart attacks or faster treatment.
Claimed by
Powerful Medical, September 2026
Drone defibrillators in service
Everdrone and Västra Götaland Region in Sweden run AED drones from six bases covering more than 300,000 residents. Everdrone says that on 19 July 2026 in Borås one of its drones brought an AED to a man in his 60s in cardiac arrest; bystanders attached it with a dispatcher's guidance, it gave a shock, and he regained a pulse at the scene. No outcome figures for the service have been published.
Achieved whena published study of the routine service reports how many patients reached by drone survive to leave hospital.
Claimed by
Everdrone, July 2026
Open
01
CPR depth from a phone
A phone app that films a person doing chest compressions on a training manikin and reports how deep and how fast each compression is, with no sensor in the manikin or on the person.
The briefCredit · unclaimed02
Test an AI CPR coach
A randomized test, on manikins, of whether people with no recent CPR training do better CPR when a voice AI coaches them than when a trained person reads them a standard dispatcher script.
The briefCredit · unclaimed
Open 01 · the brief
CPR depth from a phone
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A phone app that films a person doing chest compressions on a training manikin and reports how deep and how fast each compression is, with no sensor in the manikin or on the person.
Why it is still open
In a 2024 test, an open pose-tracking model (MediaPipe) measured compression rate well from phone video but gave "no consistently accurate data for compression depth", on nine one-minute videos from one person. Depth decides whether compressions move enough blood. EgoEMS, a 2025 dataset from the University of Virginia, estimated depth on manikins from a camera worn on the rescuer's head together with a smartwatch, and its main results are for the two combined, not for a camera alone.
Built with what exists
an ordinary phone camera; Google's MediaPipe Pose or Apple's Vision body tracking; a monocular depth model, which estimates distance from a single camera, such as Depth Anything; a training manikin that records depth, such as Laerdal's Resusci Anne QCPR, as the answer key.
First step
Borrow a recording manikin from a CPR training center, film 20 sessions with the phone on a stand, and save each manikin log under the same name as its video.
Done looks like
100 one-minute sessions from at least 20 different people, filmed from at least two angles, with the manikin's log saved beside each video. For at least 90% of sessions the app's mean depth is within 5 mm of the manikin's and its mean rate within 5 compressions a minute. Videos, logs and code are published so that anyone can rerun the score.
Open 02 · the brief
Test an AI CPR coach
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A randomized test, on manikins, of whether people with no recent CPR training do better CPR when a voice AI coaches them than when a trained person reads them a standard dispatcher script.
Why it is still open
ChatCPR was compared with dispatchers only on recordings of 12 past 911 calls, and its authors call for careful real-world testing. A search found no randomized manikin trial of an AI coach against dispatcher coaching.
Built with what exists
ChatCPR or a general AI model with a live voice interface, such as Gemini Live, the OpenAI Realtime API or an ElevenLabs voice agent; speech-to-text that copes with shouting and noise (see the Note "Fewer word errors"); manikins that record compression depth, rate and pauses; a CPR training center to host it.
First step
Write the protocol, post it on the Open Science Framework, and run five pilot volunteers in each group. Ask the host center whether its ethics board needs to review a study with volunteers.
Done looks like
At least 60 volunteers assigned at random to AI or human coaching, with the protocol posted publicly before the first volunteer starts. The results report time to first compression and the share of compressions at the depth and rate the current resuscitation guidelines set, over the first two minutes, for both groups.