Capture
Technicians and nurses capture
They photograph the printed page, cover the header and send the photo. The photo check tells them what to fix.
Clinical safety
What TraceRead is for, and what it is not
TraceRead gives an AI first read of a printed 12 lead ECG, for a qualified clinician to check. This page says who may use it, what it must not be used for, and where it falls short, with the numbers.
An AI first read is not the last word.
- A qualified doctor reviews every AI first read. Nothing reaches a patient until a doctor signs it.
- Do not rely on TraceRead alone in an emergency.
- A normal AI first read does not rule out heart disease.
- For adults, 18 and over.
Intended use
What TraceRead is
TraceRead is software. It takes a phone photo of the paper printout of a standard 12 lead resting ECG and produces an AI first read.
The AI first read holds measurements (heart rate, PR interval, QRS, QT, QTc and heart axis) and proposed patterns from a fixed list, each with a score from 0 to 100.
It is a first read for a qualified clinician to check. It does not say what is wrong with the patient, and it does not detect heart attacks. It sends nothing to a patient.
- What it is
- An AI first read of a standard 12 lead resting ECG, from a photo of the paper printout.
- Who uses it
- Technicians and nurses capture the photo. Qualified clinicians review the AI first read.
- Where
- Diagnostic centres, hospitals and HMOs (health maintenance organisations, the companies that pay for members' care) in Nigeria.
- Patients
- Adults, 18 and over.
Words used on this page
- 12 lead ECG
- The standard ECG: 12 views of the heart's electrical activity, printed as 12 wavy lines.
- Pattern
- Something TraceRead AI sees on the tracing. It is not a finding that the patient has a disease.
- Score
- How strongly TraceRead AI sees a pattern, from 0 to 100. It is not the chance that the patient has a disease. Each pattern has its own flag line, set on other ECGs so that patterns are not missed.
- PR interval
- The time from the start of the upper chambers' beat to the start of the lower chambers' beat.
- QRS
- How long the main beat of the lower chambers lasts.
- QT and QTc
- The time the lower chambers take to beat and recover. QTc is the QT corrected for heart rate.
- Heart axis
- The main direction of the heart's electrical activity, in degrees.
Users
Who may use it
Two kinds of people use TraceRead. Only one of them reads the result.
Review
Qualified clinicians review
A clinician registered with the Medical and Dental Council of Nigeria (MDCN) reviews every AI first read, keeps, edits or removes each pattern, checks every number and decides what happens next.
Adults only: 18 and over. Nothing reaches a patient until a clinician has checked and signed it.
Limits of use
What it must not be used for
These are limits of the intended use. Do not use TraceRead for any of them.
As the only check in an emergency
A patient who is unwell is treated on clinical grounds, without waiting for TraceRead. TraceRead must not be relied on alone in an emergency.
As proof that the heart is healthy
A normal AI first read does not rule out heart disease.
For children
TraceRead is for adults, 18 and over. Do not use it for children.
For anything that is not a printed 12 lead ECG
Rhythm strips on their own, printouts from a bedside monitor that are not a 12 lead ECG, photos of monitor screens, long wearable recordings (Holter) and exercise tests are not supported.
For paper printed at other settings
TraceRead reads ECGs printed at 25 mm a second and 10 mm per millivolt. Do not use it for anything else. Paper printed at 50 mm a second, or at 5 or 20 mm per millivolt, was not tested.
Rules
Four rules TraceRead follows
These are rules, not features to switch on. They apply to every AI first read.
Rule 1
A clinician reviews every AI first read
The clinician keeps, edits or removes each pattern and checks every number. Nothing reaches a patient until a clinician has checked and signed it.
Rule 2
Urgent patterns are written as possible
A pattern that may need quick attention is worded as possible and for review now. It is never worded as a confirmed finding.
- Possible very slow heart rate, review now
- Possible ST elevation pattern, review now
Rule 3
The paper is the source of truth
The rebuilt trace is drawn over the photo so the clinician can compare the two. If the rebuilt line and the paper disagree, the paper is right. When the rebuilt trace looks unreliable, TraceRead gives no AI first read and asks for a new photo.
Rule 4
No generative AI writes clinical text
Every pattern name comes from a fixed list. Every measurement is a number from fixed code. No chatbot or text generator writes any clinical wording.
Measured limits
TraceRead AI does not reliably find ST elevation
ST elevation is a rise in one part of the heartbeat line, the ST segment. In the right leads it can be a sign of a heart attack. Two tests show how far TraceRead falls short.
0.61 to 0.63
How well the nearest pattern TraceRead AI learned tells ST elevation apart (AUROC, where 1.00 is perfect and 0.50 is a coin toss), on the 176 ECGs labelled ST elevation in an ECG set from China that it never trained on. Inside PTB-XL, the same pattern scores 0.91.
Measured on ECGs labelled ST elevation in Chapman Shaoxing Ningbo (China). Source: Spike S2.
61.9%
of the same 176 ST elevation ECGs found by a rule that measures the rise, on the original 12 lead recordings (95% interval 54.3 to 69.1). On a printout layout of 3x4 with a rhythm strip it found 50.0% (42.4 to 57.6).
The rule uses the thresholds of the Fourth Universal Definition of Myocardial Infarction. It flagged about 5% and about 6% of 2,000 other ECGs (94.7% and 93.8% of them left alone). Chapman Shaoxing Ningbo. Source: Spike S4.
89.8%
found by a more sensitive setting of the same rule, but it flagged 22% of the 2,000 other ECGs. On the printout layout it found 81.2% and flagged 30%.
The setting was fixed on other ECGs before this test. Chapman Shaoxing Ningbo. Source: Spike S4.
A normal AI first read does not rule out a heart attack. A patient with chest pain or other warning signs is treated on clinical grounds, without waiting for TraceRead.
On 676 photos made by computer from Chapman ECGs, the same rule found 50.9% on the printout layout, against 50.0% on the original recordings, so the photo step is not the cause. 17 of the 676 photos (2.5%) could not be read and gave no first read (Spike S3). The rise is measured at the J point, where the main beat ends.
Measured limits
Rules for dangerous heart rates and intervals miss events
Rules that read the measurements were tested on all 21,799 PTB-XL ECGs, on 12 leads, and compared with the same rules run on a reference ECG program's numbers. They rarely raised a false alarm. They missed events, and the rule for a very slow heart rate found only 35.3%.
Very fast heart rate
150 beats a minute or more
73 of 91
Found 80.2% of the ECGs the reference program flags (95% interval 70.9 to 87.1). 1 extra alarm.
Wide and fast beats
QRS (how long the main beat lasts) of 120 ms or more, with a heart rate of 120 or more
33 of 52
Found 63.5% of the ECGs the reference program flags (95% interval 49.9 to 75.2). 9 extra alarms.
Very slow heart rate
40 beats a minute or fewer
12 of 34
Found 35.3% of the ECGs the reference program flags (95% interval 21.5 to 52.1). 1 extra alarm.
Very long QTc
500 ms or more by both ways of correcting QT for heart rate
10 of 291
Found 3.4% of the ECGs the reference program flags (95% interval 1.9 to 6.2). 4 extra alarms.
An extra alarm is an ECG the rule flagged and the reference program did not. For a very slow heart rate the beat detector counts P or T waves as beats in some slow rhythms. For a very long QTc there is no solid reference: one reference program flags 291 ECGs and another only 11. Check the heart rate and QTc on every ECG yourself.
Measured on 21,799 PTB-XL ECGs against numbers from a reference ECG program published in PTB-XL+. Source: Spike S4.
Measured limits
Rebuilding the trace can cut tall peaks
When TraceRead rebuilds a trace from a photo, the tallest peaks of the main beat can come out too short. The paper is the source of truth for heights.
On photos made by computer, about one lead in ten had its upward peaks cut to under 80% of their true height. 56 of 493 readable photos (11%) had four or more of the 12 leads cut like this.
Those photos scored 0.072 lower on the 5 broad groups of patterns, four times the drop of photos with at most one cut lead. Patterns that depend on height lost the most.
Measured on 493 readable photos made by computer of PTB-XL test ECGs. Source: Spike S3.
Measured limits
Where the numbers come from
Every number on this site was measured on public research ECGs, and on photos made by computer from them.
The ECGs come from Germany (PTB-XL), China (Chapman Shaoxing Ningbo) and Brazil (CODE-15%). The photos were made by TraceRead's own photo simulator, with tilt, light, shadow and blur.
No number on this site comes from a real phone photo of a printed ECG, and none from a patient in Nigeria. TraceRead has not been measured on ECGs from Nigerian patients, so do not assume the numbers carry over.
Regulation
Regulatory status
TraceRead is not registered with NAFDAC as a medical device.
NAFDAC is the National Agency for Food and Drug Administration and Control, which regulates medical devices in Nigeria.
Read next
More about how TraceRead works
Ask us anything about the limits
Tell us what your clinicians need to know before they use an AI first read. We answer with the measurements.