Five questions (Q1-Q5) with an additional option to comment on were drafted by P1-MUI, P2-EMTEC, P3-MUW, P5-UMCG and P8-REGIONH in the fourth reporting period and sent to international opinion leaders in the area of diabetes/endocrinology, nephrology, epidemiology and medical ethics. 21 responses were received.
Q1:
Consider the efficacy/side effect profile of drugs currently in use for DKD. What sensitivity / specificity of a decision support tool would you require to be met before incorporating the tool`s suggestions about treatment to stabilize eGFR into your clinical judgement?
I would not use such a tool because I will always prescribe multi-drug combination therapy 35%
SE 90%; SP 90% 35%
SE 90%; SP 50% 12%
SE 50%; SP 90% 6%
Other thresholds 12%
In the comments section the experts pointed out, that the drugs in focus may not only be of benefit for DKD but also cardiovascular protection. Furthermore the utility and clinical use of the tool may be affected by the patient characteristics (high risk vs. low risk).
Q2:
Assume the tool for drug response prediction has excellent SE/SP. However, the tool uses AI methodology, and you cannot rationally explain to a patient why specific recommendations are made.
Will the method be accepted by doctors?
Yes 65%
No 35%
Will the method be accepted by patients?
Yes 72%
No 28%
The majority of the experts expressed the opinion that several risk assessment tools already in use in clinical routine that also cannot be explained in detail to patients. The fact that AI is involved was not considered to make a major difference. This was elegantly summarized by one opinion leader: It is always how to explain to a patient and not what to explain.
Q3:
Assume a tool A predicts, that a specific drug will stabilize eGFR with high SE/SP. However it uses bioinformatics identified parameters and a neural network and hence the rationale of the recommendation cannot be “explained” to a patients. In contrast tool B has a lower SE/SP but you can at least try to “explain” to a patient why the prediction was made in this way.
A 47%
B 41%
None 12%
In the comments section the arguing went from absolute superiority of accuracy to physicians and patients like to have a rationale and explanations.
Q4:
Assume a tool that predicts a stabilization of eGFR by a drug with an accuracy (SE+SP/2) of 0.8 and the drug stabilizes eGFR in 50% of patients on a cohort level.
Yes 66%
No 33%
Despite this clear trend some limitations were mentioned (better use in advanced stages; also take cardiovascular risk into account).
Q5:
A decision support tool aims to bring precision medicine in DKD into clinical practice. What is the alternative to precision approaches (multiple answers possible)?
Valid answers 18 (multiple answers per person allowed).
Trial and error 10 42%
Cumulative drug prescription 42%
Other 16%
Interesting comments made were “Precision approaches are best for avoiding harm, if we go for efficacy we may risk under-treatment” and “Selection may be indicated for high cost agents but not for low cost effective agents. Cost effectiveness in general is most important in low risk populations and we need to define the threshold of use.”
About one third of the experts showed a negative attitude towards the drug decision support tool to select the best treatment options, two third showed an increasingly positive attitude with increasing SE/SP scores. The experts group was already familiar with the use of risk assessment tools in clinical routine. 65% of the participants showed little concerns even in case of a lack of rational explanations; they think that most of their patients will still accept involvement of AI (72%). Others pointed out how important reasons and explanations are for both, physicians and patients. Concerns were mentioned about cost effectiveness (low risk population), the possible loss of benefits regarding cardiovascular protection and the risk of under-treatment.
