The POET II trial of two antibiotic strategies of endocarditis is an example where you have to look beyond the primary endpoint and not rely on the authors’ conclusions
Dr. Mandrola, I agree with your bottom line: standard duration remains the default. But I think the strongest case against the authors' conclusion is in the protocol, not in the main paper.
In protocol version 1 (2018), the primary endpoint was a composite of death, embolism, and bacteremia with the same organism. The November 2023 amendment did three things after a (blinded) interim analysis: it made "days alive without antibiotics" the primary endpoint, removed relapse from the safety composite, and widened the noninferiority margin from 5 to 7.5 percentage points. In the sensitivity analysis that puts relapse back in, the upper CI bound is 5.54. That would fail the original 5-point margin.
Second, the benefits that justify a noninferiority design never materialized. Total adverse events were 150 vs. 151, antibiotic-related events 54 vs. 55, and IV-line complications 48 vs. 44. Quality of life was dropped from the endpoints, and costs and length of stay were not reported.
One caution in the other direction: "3x, up to 10x" is relative framing built on 17 events. In absolute terms it is 3.5 percentage points, an NNH of about 29, with a fragility index of 1. The signal is real and well-timed (median 28 days to relapse), but imprecise.
So the conclusion holds, and I think it is stronger than "clearly worse": more relapse, and no measured benefit to pay for it.
There is bacteremia recurrence without harm. You restart antibiotics, and the patient has a good outcome. There is recurrent bacteremia with harm, mainly death, surgery, and embolic phenomena. I still see a rationale for the study choice. Even with the upper bounds of bacteremia rates you cite, the trade-offs still strike me as fair. As a patient, I want to know whether a recurrence will cause harm. Those are the measured primary endpoints. Would I feel comfortable with the tailored strategy on myself? Yes.
Fair point, and I respect that you accept the upper bound and still choose tailored for yourself. That's a legitimate values call. But I'd add three things to the balance:
Relapse was in the original primary endpoint (protocol v1, 2018). It was removed in the 2023 amendment, and the noninferiority margin was widened from 5 to 7.5 pp. With relapse added back, the upper bound (5.54) fails the original margin.
"Relapse without harm" isn't something this trial can show. With 17 relapses, mortality was 13 vs. 11 (CI −3.1 to +4.7), which is compatible with real harm. And relapse wasn't free: each one took about 40 days of retreatment, more than the 15 days saved.
The benefit side never showed up: adverse events 150 vs. 151, antibiotic-related 54 vs. 55, IV-line complications 48 vs. 44, and quality of life was dropped as an endpoint.
So the trade-off is a measured risk for an unmeasured benefit.
Alain, can you say this in your own words instead of AI? Regardless, the last sentence is important. The risks are uncertain and the n frail. What I would balance is upper-bound mortality with gold-standard therapy. The NNH is too high. I agree.
Agree. And most of the ID doctors with whom I've discussed these results are on the same page as we are. A significantly higher relapse rates is too concerning a signal to change standard of care.
Agree, and I'd argue the case is stronger than the relapse p-value, which is fragile (p ≈ 0.05 with the protocol-specified Yates test; one fewer event and it's gone). Two more robust points:
Relapse was in the original primary endpoint (protocol v1, 2018). The 2023 amendment removed it from the safety composite and widened the noninferiority margin from 5 to 7.5 pp. With relapse added back, the upper bound (5.54) fails the original margin.
The benefits that justify a noninferiority design never appeared: adverse events 150 vs. 151, antibiotic-related 54 vs. 55, IV-line complications 48 vs. 44, and quality of life was dropped.
Curious whether your ID colleagues weighed the protocol change. It hasn't come up much in the discussion so far.
Thank you for walking through this, John! A relapse of endocarditis can mean another admission, maybe surgery on that valve. A couple more weeks of antibiotics seems like an easy trade.
Agree with the trade-off, and the data support it even more than you suggest. The "couple more weeks" cost essentially nothing measurable: adverse events were 150 vs. 151, antibiotic-related 54 vs. 55, IV-line complications 48 vs. 44. Meanwhile each relapse took about 40 days of retreatment, more than the 15 days saved, and relapses did lead to readmissions (13 vs. 4 among serious adverse events).
One nuance: surgery didn't track with relapse in this trial. Unplanned surgery was actually lower in the tailored arm (5 vs. 10). So the case rests on readmission and retreatment, not surgery, and that's strong enough.
Dr. Mandrola, I agree with your bottom line: standard duration remains the default. But I think the strongest case against the authors' conclusion is in the protocol, not in the main paper.
In protocol version 1 (2018), the primary endpoint was a composite of death, embolism, and bacteremia with the same organism. The November 2023 amendment did three things after a (blinded) interim analysis: it made "days alive without antibiotics" the primary endpoint, removed relapse from the safety composite, and widened the noninferiority margin from 5 to 7.5 percentage points. In the sensitivity analysis that puts relapse back in, the upper CI bound is 5.54. That would fail the original 5-point margin.
Second, the benefits that justify a noninferiority design never materialized. Total adverse events were 150 vs. 151, antibiotic-related events 54 vs. 55, and IV-line complications 48 vs. 44. Quality of life was dropped from the endpoints, and costs and length of stay were not reported.
One caution in the other direction: "3x, up to 10x" is relative framing built on 17 events. In absolute terms it is 3.5 percentage points, an NNH of about 29, with a fragility index of 1. The signal is real and well-timed (median 28 days to relapse), but imprecise.
So the conclusion holds, and I think it is stronger than "clearly worse": more relapse, and no measured benefit to pay for it.
Isn’t the primary objective of using antibiotics to kill the infection?
Is this approach more of the “medicate just enough to make the symptoms bearable” mentality?
There is bacteremia recurrence without harm. You restart antibiotics, and the patient has a good outcome. There is recurrent bacteremia with harm, mainly death, surgery, and embolic phenomena. I still see a rationale for the study choice. Even with the upper bounds of bacteremia rates you cite, the trade-offs still strike me as fair. As a patient, I want to know whether a recurrence will cause harm. Those are the measured primary endpoints. Would I feel comfortable with the tailored strategy on myself? Yes.
Fair point, and I respect that you accept the upper bound and still choose tailored for yourself. That's a legitimate values call. But I'd add three things to the balance:
Relapse was in the original primary endpoint (protocol v1, 2018). It was removed in the 2023 amendment, and the noninferiority margin was widened from 5 to 7.5 pp. With relapse added back, the upper bound (5.54) fails the original margin.
"Relapse without harm" isn't something this trial can show. With 17 relapses, mortality was 13 vs. 11 (CI −3.1 to +4.7), which is compatible with real harm. And relapse wasn't free: each one took about 40 days of retreatment, more than the 15 days saved.
The benefit side never showed up: adverse events 150 vs. 151, antibiotic-related 54 vs. 55, IV-line complications 48 vs. 44, and quality of life was dropped as an endpoint.
So the trade-off is a measured risk for an unmeasured benefit.
Alain, can you say this in your own words instead of AI? Regardless, the last sentence is important. The risks are uncertain and the n frail. What I would balance is upper-bound mortality with gold-standard therapy. The NNH is too high. I agree.
Agree. And most of the ID doctors with whom I've discussed these results are on the same page as we are. A significantly higher relapse rates is too concerning a signal to change standard of care.
Agree, and I'd argue the case is stronger than the relapse p-value, which is fragile (p ≈ 0.05 with the protocol-specified Yates test; one fewer event and it's gone). Two more robust points:
Relapse was in the original primary endpoint (protocol v1, 2018). The 2023 amendment removed it from the safety composite and widened the noninferiority margin from 5 to 7.5 pp. With relapse added back, the upper bound (5.54) fails the original margin.
The benefits that justify a noninferiority design never appeared: adverse events 150 vs. 151, antibiotic-related 54 vs. 55, IV-line complications 48 vs. 44, and quality of life was dropped.
Curious whether your ID colleagues weighed the protocol change. It hasn't come up much in the discussion so far.
I think you could subtitle this essay as:” When bad things happen to good data…’
Good one
Thank you for walking through this, John! A relapse of endocarditis can mean another admission, maybe surgery on that valve. A couple more weeks of antibiotics seems like an easy trade.
Agree with the trade-off, and the data support it even more than you suggest. The "couple more weeks" cost essentially nothing measurable: adverse events were 150 vs. 151, antibiotic-related 54 vs. 55, IV-line complications 48 vs. 44. Meanwhile each relapse took about 40 days of retreatment, more than the 15 days saved, and relapses did lead to readmissions (13 vs. 4 among serious adverse events).
One nuance: surgery didn't track with relapse in this trial. Unplanned surgery was actually lower in the tailored arm (5 vs. 10). So the case rests on readmission and retreatment, not surgery, and that's strong enough.