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