Ovarian cancer screening spin
Telegraph: “Screening could prevent one in five ovarian cancer deaths, study shows” here
BBC: “Ovarian cancer: Screening may cut deaths by a fifth” here
But a negative trial
“The primary analysis […] gave a[n ovarian cancer] mortality reduction over years 0–14 of 15% (95% CI –3 to 30; p=0·10) with [multi-modal screening] MMS and 11% (–7 to 27; p=0·21) with [ultrasound screening] USS.”
The UK Collaborative Trial of Ovarian Cancer Screening (UKCTOCS) recruited post-menopausal women aged between 50 and 74. 50,639 were randomly assigned to undergo ovarian cancer screening by vaginal ultrasound (USS), 50,640 to ultrasound and CA125 blood testing, so called multi-modal screening (MMS) and 101,359 to be controls. The primary outcome, death due to ovarian cancer, analysis plans and sample size were all pre-specified and followed, and the results published in last week’s Lancet (click here or ukctocs)
Neither method worked (see above). But the small favourable trend in ovarian cancer deaths, which may have been an effect of chance, only appeared after seven years. Perhaps a true effect had been swamped by women who already had cancer when screened. The authors did a second data-driven analysis and concluded:
“Although the mortality reduction was not significant in the primary analysis, we noted a significant mortality reduction with MMS when prevalent cases were excluded. We noted encouraging evidence of a mortality reduction in years 7–14, but further follow-up is needed before firm conclusions can be reached on the efficacy and cost-effectiveness of ovarian cancer screening.”
First analysis fails so, having seen the data, have another go! And Ian Jacobs and Usha Menon, the two lead authors, hold shares in the company which owns the screening algorithm! And the lack of difference in all cause mortality (RR = 0.99, slightly favouring the control arm, albeit not remotely significant) was hidden deep in the appendix (Web table 6 UKCTOCS appendix). Is this spin?
There are two possible reasons why a screening test might appear to reduce one type of cancer-related death but not alter all cause mortality.
One reason is that ovarian cancer deaths are relatively rare, so even a real effect is swamped. The 45 deaths difference between the 347 ovarian cancer deaths among controls and the 302 among the two screened groups, amounts to only 0.33 percent of the overall 13,296 deaths in the trial.
The other is that a some deaths, which would have been classified as ovarian cancer in controls, get classified as something else in the screening group. This happened. There were 11 deaths due to primary peritoneal cancer among controls but 21 in the treatment arms. Primary peritoneal cancer is the label we give ovarian cancer if the woman has already had her ovaries removed. Including these makes the primary analysis even less convincing. MMS “ovarian or primary peritoneal mortality” reduction = 11% (95% CI –8 to 26; p=0·23) and USS = 9% (–9 to 24; p=0·31) (Table 3).
Once the primary peritoneal cancers are included, even exclusion of the prevalent cases leaves a difference which is no longer statistically significant. MMS “ovarian or primary peritoneal mortality” reduction = 18% (–1 to 34; p=0·064) and USS = 17% (–3 to 33; p=0·097) weighted log rank model (post hoc) excluding prevalent cases but including primary peritoneal (table 3).
The trial showed a small reduction in ovarian cancer deaths which could have occurred by chance, and amounted to 0.33% of all deaths. After looking at the data and excluding prevalent cancers a different statistical test nominally indicated that the effect was unlikely to have occurred by chance. But they also found, as expected, that 11 women who died in the screening group of primary peritoneal cancer would have been classified as ovarian cancer deaths had they been in the control arm. So instead of nominally preventing 45 ovarian cancer deaths, the screening actually prevented only 35. Using these numbers, the difference, even with post hoc exclusion of prevalent cases, was no longer statistically significant.
The only other big trial of ovarian cancer screening (PLCO) also showed no benefit. In that trial the direction was opposite; “118 deaths caused by ovarian cancer (3.1 per 10 000 person-years) in the intervention group and 100 deaths (2.6 per 10 000 person-years) in the usual care group (mortality RR, 1.18; 95% CI, 0.82-1.71).” The inclusion of these data will reduce nominal statistical significance still further.
Would it be worthwhile if true?
The present paper reports neither the human, time sitting in clinics undergoing blood tests or with a probe up your vagina, nor the economic costs. But here’s a “back of the envelope” calculation. If 100,000 women undergo 700,000 blood tests and 700,000 vaginal scans we might possibly save 35 lives. If each test takes up just half a day (appointments, travel, counselling, waiting, testing, going home and being given results) that would amount to 959 woman-years being screened. The lives are only saved after between 7 and 14 years, so the beneficiaries will be between age 57 and 88. If optimistically each gained an average of 10 years, the programme would deliver 350 woman-years of benefit at the cost of spending 959 woman-years undergoing the screening. Doubtless more sophisticated analyses will appear, but I’d advise scepticism if anyone claims the net benefits justify the costs.
Such blatant massaging of a positive conclusion out of a negative trial depressed me. But I was cheered by this excellent blog (click here) from Cancer Research UK, who had part funded the trial. It concludes:
“We don’t think there’s enough evidence for the NHS to introduce a national screening programme at this stage.”
Well said. The NHS will save hundreds of millions of pounds by not introducing this futile screening programme; money to spend on treatments which work. Nice trial. Happy Christmas.