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