At this year’s AIMed Europe took place between 17 and 19 September in central London, Dr. Annette ten Teije, Associate Professor of Vrije Universiteit Amsterdam reminded delegates that “we don’t just have data but also medical knowledge”. Indeed, at the moment, there tend to be a heavier emphasis being put on the former rather than the latter and the importance of integrating medical data and medical knowledge is sometimes disregarded. 

Medical data and medical knowledge 

Dr. ten Teije addressed the audience during a panel discussion together with experts from different European regions. She explained medical data as a combination of clinical patient data, cohort studies, medical images, while medical domain knowledge encompasses clinical guidelines, side-effects of medications, causal relations and so on. 

Dr. ten Teije explained the differences between medical data and knowledge on stage.

She also highlighted the limitations of just focusing on either medical data or knowledge. If a machine is solely driven by data (learning), it could face challenges including insufficient data, the inability to transfer or generalize conclusions beyond the training/testing datasets, negligence of prior knowledge and the lack of explanations on how the machine could derive at an answer and so on. 

Likewise, if it is a knowledge-driven machine (reasoning), it may not be robust enough to handle the noise coming from segmented or incomplete data and its reliance on medical expertise will make it expensive because “the more knowledge you have, the higher the cost”, Dr. ten Teije said. “Knowledge used in the medical is often not related to data. Thus, if you are interested in learning, one always has to start from scratch”. 

Using knowledge to compensate for incomplete data 

As such, Dr. ten Teije proposed a hybrid system, combining both data and medical knowledge. She encouraged the AI community to design patterns which contain both learning-focused and reasoning-focused systems. For example, the use of symbolic reasoner to improve the performance of a substance learning algorithm. 

Dr. ten Teije explained her approach to integrate data and medical knowledge.

She also drafted out a table to further illustrate the use of knowledge to compensate for incomplete data, a prominent challenge facing those who are designing an AI algorithm.

SourceEnrichmentKnowledgePurpose
ConsultationAssociated diseases/
symptoms 
ICPC/SNOMED
(300,000 terms)
Investigate
comorbidity 
MedicationAssociated diseases DrugBank
(13,000 drugs) 
Investigate
comorbidity 
MedicationSide effects DrugBank/ SIDER
(140,000 interactions) 
Investigate
symptom causes
MedicationActive ingredients DrugBank/ SIDER/ Dailymed
(105,000 drugs) 
Investigate
ingredient effects 

She believes by integrating data and medical knowledge, the eventual solution that is being generated could be updated at a faster pace and they could also be immediately validated based on evidence obtained from previous research studies or treatments done on patients. “A trustworthy medical AI application requires both knowledge and data,” Dr. ten Teije added. 

Author Bio
synthetic gene empathy chinese artificial intelligence data medicine healthcare ai

Hazel Tang

A science writer with data background and an interest in the current affair, culture, and arts; a no-med from an (almost) all-med family. Follow on Twitter.