Is personalised medicine the future of healthcare?

In Steven Spielberg’s Minor­i­ty Report, police can pre­dict who will com­mit a crime before it even hap­pens. The idea that med­i­cine could pre­dict health threats has a sim­i­lar touch of sci-fi – yet it’s hap­pen­ing today. 

Canada’s Artemis Project, for exam­ple, pro­vides insights into pos­si­ble future heart attacks. In the US, the Van­der­bilt Uni­ver­si­ty Med­ical Cen­ter has been able to iden­ti­fy peo­ple at risk of devel­op­ing cer­tain auto-immune dis­eases. There are many oth­er such instances, all of which use machine learn­ing and arti­fi­cial intel­li­gence (AI) to crunch data.

The med­ical world has long employed AI in mak­ing its diag­noses. The sec­tor is an ide­al can­di­date for the tech­nol­o­gy thanks to healthcare’s data tsuna­mi, the vast major­i­ty from imag­ing tech­nol­o­gy. 

But recent years have seen rapid progress. Bare­ly a week pass­es with­out some new study sug­gest­ing AI is on a par with or even out­per­forms med­ical spe­cial­ists in their diag­noses, through its tal­ent for pat­tern recog­ni­tion. As part of a project with London’s Moor­fields Eye Hos­pi­tal, Google’s Deep­Mind can now rec­om­mend the cor­rect refer­ral deci­sion for over 50 eye dis­eases with 94% accu­ra­cy. 

Crit­i­cal­ly, AI per­forms its tasks much faster than humans. It’s also cheap­er and presents less of a bur­den on typ­i­cal­ly stretched health­care resources. And as doc­tors grap­ple with the chaos of a dynam­ic sit­u­a­tion, AI diag­no­sis will like­ly low­er the risk of med­ical mis­di­ag­no­sis, too. 

Rise of the sensors

Many of us might strug­gle with the notion of putting our health in the hands of a com­put­er. How­ev­er, “the pan­dem­ic has had a major impact on accep­tance,” reck­ons Haider Raza, lec­tur­er in com­put­er sci­ence at the Uni­ver­si­ty of Essex, who’s devel­op­ing a proof-of-con­cept sys­tem for skin can­cer detec­tion and analy­sis using smart­phone pho­tographs. 

Raza start­ed his project before the pan­dem­ic and had to con­vince peo­ple to get involved. Today, “peo­ple are sign­ing up for self-refer­ral, because Covid has shown them how so much health­care can be man­aged online.”

That’s just as well, because health­care tech is only going to play a greater role in our lives. As domes­tic med­ical devices – ever-small­er wear­ables like activ­i­ty track­ers, glu­come­ters, smart inhalers, heart rate and blood pres­sure mon­i­tors – become more com­mon­place and con­nect­ed in real time to the so-called inter­net of things, the poten­tial could be rev­o­lu­tion­ary.

Yet AI may not be the sil­ver bul­let for med­ical diag­noses that it’s some­times made out to be. 

“We’re head­ing towards some­thing good but there is still a lot of hype,” accord­ing to Maarten Van Sme­den, assis­tant pro­fes­sor of epi­demi­ol­o­gy at Uni­ver­si­ty Med­ical Cen­tre, Utrecht. 

A Uni­ver­si­ty of Birm­ing­ham study in 2019 found that most machine learn­ing algo­rithms are on a par with doc­tors in assess­ing med­ical imag­ing. How­ev­er, it con­clud­ed that of 20,530 stud­ies on dis­ease-detect­ing algo­rithms pub­lished since 2012, few­er than 1% were rig­or­ous enough to be includ­ed in its study in the first place. 

“If you need an AI mod­el to make a diag­no­sis, ask why, because it prob­a­bly means it’s hard to make a diag­no­sis. The fact is that AI needs many high-qual­i­ty data points to dis­tin­guish between those with a dis­ease and those with­out, so it becomes a cir­cu­lar prob­lem,” says Van Sme­den.

Data challenges

Shang-Ming Zhou, pro­fes­sor of e‑health at the Uni­ver­si­ty of Ply­mouth, thinks we’re only start­ing to grap­ple with the many issues revealed by the use of AI in med­i­cine. The data sets are often small­er than might be hoped, due to issues around pri­va­cy, patient con­fi­den­tial­i­ty and data own­er­ship. What’s more, var­i­ous data sources may be incon­sis­tent depend­ing on how they’re pro­duced, poten­tial­ly build­ing prob­lems into any AI mod­el. 

Data, like doc­tors, comes with its own bias­es. An algo­rithm trained on Cau­casian pop­u­la­tion data may pro­vide mis­di­ag­noses for oth­er eth­nic­i­ties. Some dis­eases, the likes of sick­le cell or Tay-Sachs, are shaped by eth­nic­i­ty, oth­ers by geog­ra­phy.

“You wouldn’t expect biased data to deliv­er fair pre­dic­tions, but cur­rent AI devel­op­ment in health­care isn’t address­ing that kind of eth­i­cal issue,” says Zhou. Then there’s the chal­lenge of reg­u­la­tion, with no cur­rent legal frame­work for data pro­tec­tion in pri­vate health­care research. 

“The oth­er chal­lenge is that the cur­rent AI mod­el can­not gen­er­alise to a new pop­u­la­tion of patients or con­sid­er that health­care prac­tices evolve over time,” he adds. 

That’s why talk of AI replac­ing human doc­tors – whose insights are born of expe­ri­ence and patient inter­ac­tions – still seems far-fetched. “AI is a pow­er­ful tool, but a tool nonethe­less,” says Van Sme­den. 

The use of AI diag­no­sis by med­ical pro­fes­sion­als could remain in the back­ground, large­ly unknown to patients. But trans­paren­cy will be key, argues Zhou, who’s cur­rent­ly research­ing patient atti­tudes to AI. That’s a prob­lem, because AI’s “think­ing” is opaque in reach­ing its con­clu­sions, unac­cept­ably so for health­care.  This can allow unno­ticed errors to become sys­temic faults; a bad mod­el can end up harm­ing patients. 

“The con­clu­sions that AI reach­es have to be ful­ly explain­able and ful­ly inter­pretable,” stress­es Saurabh Johri, chief sci­ence offi­cer at Baby­lon, a dig­i­tal health­care spe­cial­ist. “A sig­nif­i­cant pro­por­tion of the pop­u­lace is informed about data and open to being bet­ter informed as a result of its use. But what they still want to know is the val­ue of that data and they can’t with­out trans­paren­cy.”

A matter of trust 

So why don’t patients trust AI? It’s not because they think it will fail to give a bet­ter diag­no­sis, says Zhou. Rather, it’s “because the per­cep­tion is that it can only pro­vide stan­dard­ised prac­tice and treat­ment — it doesn’t address the med­ical needs of the indi­vid­ual”. Each per­son has a unique pro­file, but cur­rent AI is only suit­ed for the “aver­age” patient. 

Oth­er big ques­tions must also be resolved. For exam­ple, does the use of AI in diag­noses chal­lenge the author­i­ty of the clin­i­cian? Are its diag­noses aimed sole­ly at extend­ing life, per­haps ignor­ing a patient’s wish­es to instead min­imise suf­fer­ing? Does it under­mine the tra­di­tion­al doc­tor-patient deci­sion-mak­ing process, and does this force doc­tors to align their stan­dards with that of the algo­rithm, or to defer their deci­sions to it?

But progress is being made. The impact of machine learn­ing in health­care will like­ly be pro­found, not just in treat­ment after diag­no­sis, but in head­ing off dis­ease in the first place.

“If you want to diag­nose a dis­ease, of course that’s fine, but if you want to treat a dis­ease you need to under­stand how it devel­ops. And nowa­days we can mea­sure what genes are expressed at the sin­gle-cell lev­el and use machine learn­ing to pin­point what’s impor­tant, what is dri­ving a dis­ease,” explains Ziv Bar-Joseph, pro­fes­sor of com­pu­ta­tion­al biol­o­gy at Carnegie Mel­lon School of Com­put­er Sci­ence. 

“We need machine learn­ing for the DNA analy­sis that will allow us to under­stand which dis­eases peo­ple are pre­dis­posed to in the first place,” he notes, adding that if we can deploy some of the AI-dri­ven tests cur­rent­ly in devel­op­ment, “it will be amaz­ing.”