AI in medical decision-making: a question of ethics?

Arti­fi­cial intel­li­gence to detect breast can­cer¹ or pro­state can­cer², algorithms that cal­cu­late our physiolo­gic­al age to pre­dict age­ing³ or con­ver­sa­tion­al agents to mon­it­or our men­tal health⁴… Arti­fi­cial intel­li­gence tools are gradu­ally becom­ing part of med­ic­al prac­tice. The focus is on pre­dict­ive and per­son­al­ised medi­cine, as well as thera­peut­ic decision-mak­ing by the med­ic­al pro­fes­sion. But how do patients per­ceive this rela­tion­ship between doc­tors and AI? How do prac­ti­tion­ers really inter­act with the technology?

These are the ques­tions posed by Dami­en Lac­roux, philo­soph­er of sci­ence and research­er at the UNESCO Chair in the Eth­ics of the Liv­ing and the Arti­fi­cial. “Dur­ing my inter­views, I noticed that patients ima­gine a par­tic­u­lar rela­tion­ship between doc­tors and AI,” explains the research­er, who spe­cial­ises in the integ­ra­tion of algorithms in can­cer­o­logy. “We tend to believe that human spe­cial­ists in onco­logy delib­er­ate on our case before mak­ing a decision, and that the tech­no­logy inter­venes at a later stage to val­id­ate the delib­er­a­tion,” he explains. But is this really the case?

To find out, Dami­en Lac­roux spoke to sci­ent­ists from the MIRACLE⁵ pro­ject. This ambi­tiously named European study was launched in 2021 and brings togeth­er labor­at­or­ies in Italy, Spain, Ger­many and France. The aim is to identi­fy risk of recur­rence in lung can­cer patients, using an algorithm to help med­ic­al decision-mak­ing. To achieve this, research­ers are train­ing AI (machine learn­ing) in a super­vised man­ner. The algorithm is “fed” with data from a cohort of patients for whom the exist­ence or absence of recur­rence is known. The data inges­ted is of three types: clinico-patho­lo­gic­al data (such as the patient’s sex, the his­tory of their dis­ease or the treat­ments they may have under­gone); med­ic­al ima­ging data; and finally, omics data, i.e. a mass of inform­a­tion relat­ing to molecu­lar bio­logy (DNA or RNA from tumours).

Using a cohort of 220 patients, the sci­ent­ists feed the algorithm with all the data col­lec­ted, as well as inform­a­tion on wheth­er or not a recur­rence happened – and how long before it does. “Then we let the algorithm do its work! This involves an unima­gin­able amount of data, which is impossible for humans to pro­cess on their own,” explains Dami­en Lac­roux. “Today, the pro­ject is behind sched­ule, and we’ve only just fin­ished col­lect­ing data from the first cohort. We still have to start train­ing the algorithm with this data and then recruit a second cohort to val­id­ate its train­ing.” So, we’ll have to wait a little longer before we see the MIRACLE pro­ject in action.

But the way it works imme­di­ately raises an eth­ic­al issue, which was poin­ted out by the research­ers inter­viewed by Dami­en Lac­roux. “At the start of train­ing, bioin­form­aticians man­age to slice up the data­sets and asso­ci­ate the AI res­ults with this or that input factor. But gradu­ally, the data increases and it becomes a black box.” This grow­ing volume of data makes the mod­els used to refine pre­dic­tions more com­plex. And therein lies the para­dox: as the amount of data increases, the algorithm’s mar­gin for error decreases. The AI is there­fore more effi­cient, but the way it works is less clear to prac­ti­tion­ers. How can they explain the decisions made by AI to patients, or guar­an­tee the absence of bias, if they them­selves are not famil­i­ar with its inner workings?

In the field of onco­logy, decision trees are often used to help doc­tors jus­ti­fy their clin­ic­al reas­on­ing. How­ever, the integ­ra­tion of algorithmic scores into these pro­cesses can con­flict with the need for trans­par­ency on the part of doc­tors, who some­times struggle to under­stand what input data has led the AI to estim­ate the prob­ab­il­ity of recur­rence. “Even if we man­aged to decipher every intern­al cal­cu­la­tion in the algorithm, the res­ult would be so math­em­at­ic­ally com­plex that doc­tors would not be able to inter­pret it or use it in their clin­ic­al prac­tice,” explains a Ger­man bioin­form­atician work­ing on the MIRACLE pro­ject, inter­viewed⁶ by Dami­en Lac­roux in his forth­com­ing study.

This also affects the notion of informed patient con­sent. “The doc­tor is obliged to provide suf­fi­cient inform­a­tion to enable the patient to accept or refuse treat­ment. How­ever, if the prac­ti­tion­er is not fully informed, this poses an eth­ic­al prob­lem,” adds the philo­soph­er. And yet, as Dami­en Lac­roux points out in his study: “Molecu­lar bio­logy has iden­ti­fied the need to take into account thou­sands of pieces of patient omic data as an essen­tial means of mak­ing pro­gress in onco­logy.” AI would there­fore enable bet­ter man­age­ment of the poten­tial evol­u­tion of the dis­ease, by refin­ing the pro­posed treat­ments… at the expense of trust between doc­tors and patients.

Wheth­er AI is integ­rated into the med­ic­al delib­er­a­tion pro­cess (what Dami­en Lac­roux calls “ana­lyt­ic­al deliberation”in his art­icle) or wheth­er it is totally out­side the decision-mak­ing pro­cess and only inter­venes as a final con­sulta­tion (“syn­thet­ic delib­er­a­tion”), there must be total trans­par­ency as far as patients are con­cerned. The main fear raised by the research­er dur­ing group inter­views with patients⁷ remains that “the machine” will make the dia­gnos­is without human inter­ven­tion. “But this is not at all the case today,” reas­sures Dami­en Lacroux.

These algorithmic scores, which pro­pose a prob­ab­il­ity of can­cer recur­rence based on patient data, also raise oth­er ques­tions spe­cif­ic to pre­dict­ive medi­cine: when are we really cured? Can we really be free of the dis­ease when uncer­tainty per­sists, and we live in con­stant anti­cip­a­tion of a pos­sible recur­rence? These ques­tions, like so many oth­ers, have yet to be answered.

Sophie Podevin