Tempus AI Inc. has received United States Food and Drug Administration 510(k) clearance for Tempus ECG-PH, software that uses machine learning to identify signals associated with pulmonary hypertension from a standard resting 12-lead ECG. The clearance gives Tempus its third FDA-cleared cardiovascular artificial-intelligence product and expands the company from oncology-heavy diagnostics and data services into another area where routine clinical data may be converted into additional diagnostic information.
The most interesting part of the clearance is not simply that artificial intelligence can identify pulmonary-hypertension signals. The FDA summary shows a more nuanced performance profile: sensitivity of 87.9%, specificity of 71.5%, positive predictive value of 15.5% and negative predictive value of 99.0% under an assumed disease prevalence of 5.6% in the intended-use population.
Those numbers help define where ECG-PH may fit clinically. It is not a diagnostic replacement. It may instead become a relatively inexpensive screening and referral tool layered onto an ECG that a symptomatic patient is already receiving.
What exactly is Tempus ECG-PH cleared to do?
The device is intended for patients aged 40 years and older with cardiovascular symptoms such as breathlessness, fatigue, chest pain or oedema who do not already have a diagnosis of pulmonary hypertension. It analyses a resting, non-ambulatory 12-lead ECG and returns a binary output indicating whether elevated pulmonary-hypertension risk has been detected.
The FDA label places important limits around that output. ECG-PH is not a stand-alone diagnostic, cannot be used to direct therapy, should not trigger invasive right-heart catheterisation by itself and is not intended for serial monitoring. It should not be used on paced rhythms, and repeated testing of the same patient within 12 months has not been validated.
That makes the technology easier to understand as a clinical triage layer. A physician sees a symptomatic patient, a standard ECG is performed, and software extracts a risk signal that may otherwise be invisible in routine interpretation. The result can then contribute to the decision about whether more specific investigation is warranted.
Why do the sensitivity, specificity and predictive values matter?
Tempus trained the locked machine-learning model using more than 530,000 ECGs. Clinical performance was subsequently evaluated in an independent dataset of more than 1,000 ECGs drawn from three geographically distinct United States sites. The validation population had a median age of 69, with 53% women, and included materially greater racial diversity than the training dataset.
Sensitivity of 87.9% means the system detected a large majority of patients meeting the specified pulmonary-hypertension criteria. Specificity of 71.5% means false-positive results remain a meaningful consideration. The 99.0% negative predictive value is potentially useful because a negative output was highly associated with absence of disease under the assumed prevalence used in the analysis.
The 15.5% positive predictive value is equally important. A positive ECG-PH result does not mean a patient has pulmonary hypertension. Many patients flagged as elevated risk will not ultimately receive the diagnosis, which is why the FDA explicitly prevents the software from being used as the sole basis for invasive testing or treatment.
Commercial adoption will therefore depend on whether clinicians see value in the additional referrals generated by positive alerts. A screening technology can tolerate modest positive predictive value if the disease is serious, diagnosis is commonly delayed and the initial test is inexpensive, but it must still avoid overwhelming downstream services with unnecessary examinations.
Could pulmonary hypertension be particularly suited to AI-enabled ECG screening?
Pulmonary hypertension can be difficult to identify early because symptoms such as fatigue and shortness of breath overlap with many cardiac and pulmonary disorders. Definitive haemodynamic assessment may require right-heart catheterisation, while echocardiography is commonly used to estimate probability and guide further evaluation.
An ECG is far easier to obtain and is already embedded across primary care, emergency departments, cardiology and hospital medicine. Tempus is therefore not asking healthcare systems to install a new imaging modality simply to generate the initial signal. Its software can theoretically extract additional information from an existing workflow.
A 2026 systematic review and meta-analysis covering five studies and almost 98,000 participants reported pooled sensitivity of 0.83, specificity of 0.80 and an area under the curve of 0.88 for AI-enhanced ECG approaches to pulmonary-hypertension detection. That does not validate Tempus’ specific product, but it provides broader evidence that ECG waveforms contain machine-readable information potentially useful for the condition.
Why could workflow integration matter more than the algorithm itself?
Medical AI products often demonstrate respectable retrospective performance yet struggle to generate revenue because hospitals must decide who receives the test, where results appear, how clinicians respond and who pays for the workflow. Tempus ECG-PH has no dedicated user interface and is designed to exchange data and results through existing medical systems, including electronic-health-record and hospital-information environments.
That architecture could reduce adoption friction. If the algorithm runs automatically on eligible ECGs and returns a result in a system clinicians already use, the behavioural change required may be smaller than for a standalone application.
The business model still needs to prove itself. Hospitals will ask whether using the software identifies clinically important cases earlier, whether it alters referral patterns, whether downstream testing is manageable and whether reimbursement or broader system economics justify deployment. Clearance answers the regulatory question but not the health-economic one.
How important is cardiovascular AI to Tempus AI’s wider strategy?
Tempus reported second-quarter 2026 revenue of $382.5 million, up 22% year over year, with diagnostics contributing $289.3 million and data and applications producing $93.2 million. The company also reported $820.7 million of cash and marketable securities and raised its full-year revenue guidance to approximately $1.595 billion to $1.605 billion.
Those numbers show that ECG-PH does not need to carry the company economically in the near term. Tempus’ larger businesses remain oncology diagnostics and data licensing. That may actually benefit the cardiovascular strategy because the company can build adoption over time without relying on a single AI device to support the organisation.
The third cardiovascular clearance also creates portfolio effects. Health systems may be more willing to integrate a software infrastructure capable of supporting multiple ECG-derived algorithms than one serving a single narrow indication. Tempus already has cleared products targeting atrial fibrillation and low ejection fraction, allowing it to build a broader cardiovascular screening proposition from the same fundamental data source.
What does the clearance say about the FDA’s changing approach to medical AI?
One detail in the clearance deserves attention: the FDA authorised a Predetermined Change Control Plan for ECG-PH. Such plans can permit specified changes to an AI-enabled device without requiring a new 510(k) submission each time, provided modifications remain within the authorised framework.
That matters because machine-learning products create a regulatory challenge traditional static devices do not. Developers want to improve algorithms as data accumulate, while regulators need assurance that modifications do not introduce new safety or performance problems. The FDA is simultaneously seeking feedback on how it should regulate generative-AI-enabled medical devices, including premarket assessment, post-market monitoring and risk evaluation, showing that the broader framework remains under active development.
Tempus ECG-PH therefore sits at an interesting intersection. It is a cleared, locked machine-learning product operating under a defined clinical indication, not an open-ended generative AI system. Yet its regulatory structure points toward a future in which software devices are expected to evolve under controlled plans rather than remain permanently frozen.
For Tempus, the commercial question is whether that sophistication translates into everyday clinical value. An 87.9% sensitivity figure attracts attention, but the most important metric may ultimately be simpler: whether hospitals using ECG-PH find patients earlier without creating an unmanageable wave of false-positive follow-up.
