In­cor­po­rat­ing Ex­ter­nal Da­ta in­to Clin­i­cal Tri­als: Com­par­ing Dig­i­tal Twins to Ex­ter­nal Con­trol Arms

Most drug de­vel­op­ment pro­fes­sion­als are fa­mil­iar with the nerve-rack­ing wait for the read-out of a large tri­al. If it’s neg­a­tive, is the in­ves­ti­ga­tion­al ther­a­py in­ef­fec­tive? Or could the fail­ure re­sult from an un­fore­seen flaw in the de­sign or ex­e­cu­tion of the pro­to­col, rather than a lack of ef­fi­ca­cy? The team could spend weeks an­a­lyz­ing da­ta, but a de­fin­i­tive an­swer may be elu­sive due to in­suf­fi­cient pow­er for such analy­ses in the al­ready com­plet­ed tri­al. These prob­lems are on­ly made worse if the tri­al had low­er en­roll­ment, or high­er dropout than ex­pect­ed due to an unan­tic­i­pat­ed event like COVID-19. And if a tri­al is neg­a­tive, the next one is like­ly to be larg­er and more cost­ly — if it hap­pens at all.

  • The phar­ma­ceu­ti­cal in­dus­try spends over $80 bil­lion on R&D each year.1
  • It takes 12 to 15 years on av­er­age for an ex­per­i­men­tal drug to trav­el from the lab to U.S. pa­tients.2
  • A phase III clin­i­cal tri­al can re­quire up to 3,000 pa­tient vol­un­teers and on­ly ap­prox­i­mate­ly 25% to 30% of drugs move to the next phase.3

Spon­sors fre­quent­ly use ex­ter­nal con­trol arms to short­en time­lines and cut costs for proof-of-con­cept stud­ies. Un­for­tu­nate­ly, this opens the door to con­founders — dif­fer­ences be­tween the pa­tients in the ex­ter­nal con­trol arm and those in the treat­ment arm that make it im­pos­si­ble to at­tribute dif­fer­ences in out­comes to the ef­fect of the treat­ment. As a re­sult, tri­als with ex­ter­nal con­trol arms are ef­fi­cient but un­re­li­able.

The avail­abil­i­ty of large his­tor­i­cal datasets of lon­gi­tu­di­nal pa­tient in­for­ma­tion and the rapid de­vel­op­ment of ar­ti­fi­cial in­tel­li­gence (AI) tech­nolo­gies mean that clin­i­cal tri­als don’t have to re­main stuck in this sta­tus quo. It’s pos­si­ble to have the best of both worlds: the re­li­a­bil­i­ty of a ran­dom­ized con­trolled tri­al cou­pled with the ef­fi­cien­cy of an ex­ter­nal con­trol arm. The in­no­va­tion that makes this pos­si­ble is called a Dig­i­tal Twin.


A Dig­i­tal Twin is a com­pre­hen­sive, lon­gi­tu­di­nal clin­i­cal record cre­at­ed us­ing the base­line da­ta col­lect­ed from a pa­tient — be­fore they re­ceive their first treat­ment — that pre­dicts how that pa­tient would like­ly evolve over the course of the tri­al if they were to be giv­en a place­bo. That is, a Dig­i­tal Twin is like a sim­u­lat­ed con­trol group for a par­tic­u­lar pa­tient.

Dig­i­tal Twins are treat­ed as co­vari­ates op­ti­mized to ex­plain vari­abil­i­ty in the out­come. By us­ing pre-spec­i­fied co­vari­ate ad­just­ment — ef­fec­tive­ly, com­par­ing pre­dict­ed place­bo out­comes to ac­tu­al place­bo out­comes and cor­rect­ing any bias — Dig­i­tal Twins can be in­cor­po­rat­ed in­to ran­dom­ized con­trolled tri­als to im­prove pow­er and ef­fi­cien­cy with­out in­tro­duc­ing bias.4

Dig­i­tal Twins are not Syn­thet­ic Con­trols. Syn­thet­ic Con­trol Arms (SCAs) add pa­tients who were not in the orig­i­nal pa­tient sam­ple of the tri­al. Be­cause SCAs can in­crease Type I er­ror, use cas­es are lim­it­ed based on reg­u­la­to­ry guid­ance.5 Dig­i­tal Twins add in­for­ma­tion about pa­tients al­ready in the tri­al. Be­cause these pre­dict­ed out­comes are treat­ed as co­vari­ates, they do not in­tro­duce bias, and can be ap­plied to all stages of clin­i­cal de­vel­op­ment.

Read Un­learn.AI’s whitepa­per, In­cor­po­rat­ing Ex­ter­nal Con­trol Arms in­to Clin­i­cal Tri­als, to learn about SCAs and Dig­i­tal Twins, their re­spec­tive sta­tis­ti­cal tech­niques, and ef­fects on tri­al op­er­at­ing char­ac­ter­is­tics. Al­though both ap­proach­es in­crease sta­tis­ti­cal pow­er, un­like SCAs, Dig­i­tal Twins are ro­bust to known and un­known founders.



Don’t miss out on our up­com­ing we­bi­nar! Reg­is­ter be­low.