Applying machine learning in our hospital system

What if we could har­ness data cap­tured every day in the pro­vi­sion of care and pro­vide valu­able insights that sup­port clin­i­cal deci­sions, for exam­ple, a sug­gest­ed treat­ment and care plan tai­lored to opti­mise out­comes for a spe­cif­ic patient? 

Machine Learn­ing (ML) is used for pro­cess­ing large and com­plex data to build use­ful pre­dic­tive mod­els that have the poten­tial to rev­o­lu­tionise health­care through per­son­alised and pre­ci­sion med­i­cine, aug­ment­ing deci­sion mak­ing and remov­ing cog­ni­tive load for clinicians. 

Work­ing on a range of clin­i­cian led projects with part­ners from across the health sys­tem, we are cre­at­ing and/​or test­ing ML mod­els to improve patient care with the aim of reduce ramp­ing in our hos­pi­tal’s Emer­gency Departments. 

This includes:

• Esti­mat­ing a patient’s Length of Stay based on a clin­i­cal spe­cial­i­ty, reduc­ing admin­is­tra­tive bur­den and sup­port­ing reduc­tion in length of stay.
• Pro­vid­ing diag­no­sis tools that use stan­dard tests and obser­va­tions to help aug­ment deci­sion mak­ing in gen­er­al med­i­cine wards with a view to mov­ing patients to appro­pri­ate care set­tings faster.
• Iden­ti­fy­ing patients who are at risk of a new admis­sion with­in a month of dis­charge from hos­pi­tal to pro­vide oppor­tu­ni­ties for interventions.
• Sup­port­ing effi­cien­cies in the gen­er­a­tion of dis­charge sum­maries, to improve con­ti­nu­ity of care and han­dover to Gen­er­al Prac­tice (reduc­ing admissions). 

How­ev­er, there are a num­ber of chal­lenges that need to be addressed before ML can be wide­ly imple­ment­ed in hos­pi­tal systems. 

One of the biggest chal­lenges is the lack of stan­dard­ised data. Health data is com­plex and vari­able, and it can be dif­fi­cult to com­bine data from dif­fer­ent sources and from dif­fer­ent health sys­tems in a way that is com­pat­i­ble with ML algorithms. 

Impor­tant diag­nos­tic infor­ma­tion, such as notes describ­ing the treat­ment of a patient are stored in unstruc­tured text doc­u­ments which makes access­ing the infor­ma­tion even more challenging. 

It can be dif­fi­cult to answer or even pose new ques­tions if it requires the cre­ation of a new bespoke dataset. 

To address these chal­lenges, we are work­ing to com­bine data from core sys­tems, such as sur­gi­cal and ambu­lance data, to cre­ate well doc­u­ment­ed, curat­ed and reusable data sets that are organ­ised, cleansed and val­i­dat­ed in col­lab­o­ra­tion with SA Health’s clin­i­cians and oper­a­tional staff. These data sets, once anonymised, can be used to cre­ate ML mod­els more quick­ly and eas­i­ly that will help facil­i­tate a data dri­ven approach to improv­ing health care across the state. 

The sys­tem­at­ic adop­tion of ML in health­care is still in its ear­ly stages, but the poten­tial ben­e­fits are enor­mous. By address­ing the chal­lenges of apply­ing ML in health­care, we can improve the qual­i­ty, effi­cien­cy, and afford­abil­i­ty of health­care for everyone. 

To know more, reach out to our Clin­i­cal Infor­mat­ics team at ceih@​sa.​gov.​au.