P129: Artificial Intelligence/Machine Learning advances in Real World Data Analytics
Poster Presenter
Jacqueline Vanderpuye-Orgle
Vice President
Parexel United States
Objectives
Real world data (RWD) in the form of electronic health records (EHRs) has gained popularity recently. However, its synthesis remains challenging as >80% of the information is noted as free text. The current paper focuses on advancements in the use of ArtificiaI Intelligence/Machine Learning for RWD.
Method
Targeted literature search was conducted to identify publications that discuss the use of artificial intelligence/machine learning techniques for RWD analytics. After reviewing the identified publications, evidence on latest and unique techniques was assimilated.
Results
Analyses of EHRs have evolved over time, with traditional statistical techniques giving way to recent advances in machine learning (ML). The landscape is shifting from techniques like support vector machines and random forests to the use of more innovative methods such as deep learning and neural networks for data mining and reliable predictive modelling from EHR data. The number of potential predictors from data generated by EHR may reach hundreds, in case free-text notes from healthcare providers are included. Modeling techniques which take complex nonlinear interactions among variables in line with the individual’s medical history are required for efficient learning. A US Food and Drug Administration (FDA) funded project led by Mark Van der Laan and Susan Gruber explored the use of targeted learning (TL) comprising of targeted minimum loss-based estimation (TMLE) and super learning (SL) for regulatory decision making. The study showed that TL techniques provided efficient estimates and valid inference, with TMLE supporting estimation of bias and SL supporting mitigation of bias. TL was instrumental in extracting reliable insights from RWD, a situation which is often complicated due to absence of treatment randomization, intercurrent events, and loss to follow-up. TL’s approach to learning from data provided transparency, thereby presenting a method for RWD to be incorporated in regulatory decision making.
Conclusion
The advancement of real-world data (RWD) analytical techniques holds immense potential for unlocking unique insights from patient data. Maintaining rich, standardized patient data, like the Medical Information Mart for Intensive Care (MIMIC) database, enables collaboration among academia, industry leaders, and the pharmaceutical industry. MIMIC has been used to assess techniques like gradient boosting and deep learning, benefiting critical care management and public health.
Developments in analytic techniques should be coupled with advances in data storage, analysis, and acquisition facilitate ML algorithms and frameworks to continuously learn from additional data sources and drive optimal drug development. However, there is a dearth of specific guidance on credible and acceptable use of ML techniques for predictive RWD analytics. Regulatory oversight is crucial to ensure safety, quality, and validity of AI/ML insights.
