How probable is it that someone will suffer a heart attack, develop diabetes or mental health problems in the next 20 years? A new AI model, deve\u00adloped by the Euro\u00adpean Mole\u00adcular Biology Labo\u00adra\u00adtory (EMBL) and the German Cancer Rese\u00adarch Centre (DKFZ), among others, promises predic\u00adtions for over 1,000 dise\u00adases \u2013 based on real-world data.<\/p>\n
From indi\u00advi\u00addual dise\u00adases to the big picture<\/b><\/p>\n
Previous models have mostly focused on a single disease \u2013 such as cardio\u00advas\u00adcular risk or genetic predis\u00adpo\u00adsi\u00adtions for cancer. Delphi-2M thinks bigger: it has been trained with data sets from more than two million people, combines genetic, clinical and demo\u00adgra\u00adphic infor\u00adma\u00adtion, and uses this to calcu\u00adlate proba\u00adbi\u00adli\u00adties for a wide variety of clinical pictures. At its core, it is a so-called Gene\u00adra\u00adtive Pre-trained Trans\u00adformer (GPT) model \u2013 an algo\u00adrithm that reco\u00adg\u00adnises patterns and makes predic\u00adtions based on them.<\/p>\n
Delphi-2M performed remar\u00adkably well in tests: for dise\u00adases with clear progres\u00adsion patterns, such as certain types of cancer or heart attacks, the model achieved higher accu\u00adracy than specia\u00adlised indi\u00advi\u00addual models in some cases. It was more diffi\u00adcult with psych\u00adia\u00adtric disor\u00adders, pregnancy compli\u00adca\u00adtions or rare condi\u00adtions, where the progres\u00adsion is more complex and the data available is more scarce.<\/p>\n
Oppor\u00adtu\u00adni\u00adties and pitfalls<\/b><\/p>\n
Experts are impressed \u2013 but also cautious. Prof. Robert Ranisch (Univer\u00adsity of Potsdam) sees Delphi-2M as \u2018an impres\u00adsive example of the poten\u00adtial of gene\u00adra\u00adtive AI in health rese\u00adarch\u2019. But he warns that bias, discri\u00admi\u00adna\u00adtion and the respon\u00adsible hand\u00adling of sensi\u00adtive health data remain key challenges.<\/p>\n
The ethical dimen\u00adsion is also under discus\u00adsion: no one should be confronted with a personal risk analysis without having given their consent. The right not to know must be preserved, as must data protection.<\/p>\n
But it\u2019s not just about predic\u00adtions. As Carsten Marr from the Helm\u00adholtz Centre in Munich points out: \u2018What is parti\u00adcu\u00adlarly exci\u00adting is the disco\u00advery of previously unknown corre\u00adla\u00adtions between dise\u00adases \u2013 connec\u00adtions that we might other\u00adwise never have recognised.\u2019<\/p>\n
AI oracle or tool?<\/b><\/p>\n
The average AUC (area under the curve, a measure of the perfor\u00admance of a clas\u00adsi\u00adfi\u00adca\u00adtion model) of 0.76 shows that For indi\u00advi\u00addual pati\u00adents, Delphi-2M is not yet a precise oracle. But as a rese\u00adarch tool and basis for preven\u00adtion stra\u00adte\u00adgies, it has enormous potential.<\/p>\n
Ulti\u00adm\u00adately, Delphi-2M remains what its name promises \u2013 a tool that allows us to look into the future, but not a machine of destiny. How we deal with these insights is less a ques\u00adtion of tech\u00adno\u00adlogy than one of ethics.<\/p>\n
Or, as Ranisch puts it: \u2018Such predic\u00adtions are not verdicts of fate \u2013 but valuable poin\u00adters for preven\u00adtion and therapy.\u2019<\/p>\n