Discounting risk prediction models – a reply

I thank Drs Wong and Bedford for their letter. After personal communication, I can confirm that we agree that long-term survival predictions (years) provide more valuable information for patients considering surgical treatment than mortality predictions of short duration (weeks). Short-term survival prediction has been useful for distributing the intensity of peri-operative care among surgical patients. However, as I remarked at the end of my paper, prioritising patients most likely to die in the next month might not be the most effective investment of limited resources, as more life-years might be saved by trying to prevent relatively infrequent morbidity and death in patients with long median life expectancies 1. We disagree on whether average population survival data can be modified to accurately project years of postoperative survival: I say that it can, Drs Wong and Bedford say that it cannot. I have used a calculator, which modifies general population survival, to generate accurate projected survival after repair of intact abdominal aneurysms, after cardiac surgery and after resection of colorectal cancer (I win) 2-6. The same calculator shows that survival after oesophagectomy is very different to that for a matched general population whose mortality is temporarily increased by surgery (Drs Wong and Bedford win) 7. Whether I win or Drs Wong and Bedford win, the comparison of observed postoperative survival with that expected for a matched general population is informative, whether through their agreement or disagreement. What dictates the white space between survival curves and what can we do about it? Oesophagectomy serves as an example that a few months follow-up is inadequate to characterise postoperative survival (adjusted general population survival did match this early period). There were a couple of minor points that Drs Wong and Bedford made that I would rebuff. Their 1971 reference of exponential modelling for population survival is – as they wrote – crude; it does not represent the methods that I use. They also wrote that I would not use factors other than year of birth, age and sex to generate survival curves, which is also incorrect. Readers are welcome to download copies of my calculator from https://www.dropbox.com/sh/i57xhrtyop0pq31/AACEvMTOPRzdMEQnhLbCL2G8a?dl=0 or email me. The NELA model for mortality within 30 days of emergency laparotomy will discriminate between individuals better than most peri-operative models, as mortality is relatively high and the signal relatively strong. I think that 39,000 patients would be most useful if survival were observed over subsequent years, rather than limited to 30 days. I am pessimistic about the value of the PQIP aim of studying 70,000 patients if follow-up is similarly limited to 30 days (although I understand their intent is to follow-up to potentially 20 years with NHS Digital/ONS mortality tracking). My colleague, Dr Tom Clark, and I are working on how we might modify general population survival to fit observed survival following admission to intensive care in our hospital (Fig. 1). Early mortality is greatest for non-surgical admissions, then admission after emergency surgery and least after elective surgery. We think that it will be difficult to modify general population survival to match the first two of these survival curves: we make the attempt, ‘Not because it is easy, but because it is hard’! We are not content to conclude ‘These patients are different to the general population’, we wish to know ‘the how’ and we wish to know ‘the why’. In turn, this will lead to ‘the what’ are we going to do about it?