Machine Learning (ML)/Artificial Intelligence (AI)

AI is about optimization in a stochastic environment, so it will help processes become more efficient, eliminate unnecessary effort, and provide more accurate and timely services to the end-user. Examples of AI applications may include innovative patient treatments (e.g., drug discovery, clinical trials, genomics), personalized care at scale (e.g., precision health, decision support, virtual care, customer services), elevated provider and beneficiary experience interacting with the government, and improved program integrity (e.g., fraud, waste, and abuse). The government will gain cost savings opportunities through improved business efficiency, security, and integrity.

eSimplicity implements three principal AI-enabled pipelines for Discovery, Prediction, and Recommendation that are all supported by the Data Storage and Query, Feature Extraction, Interpretability, Model Orchestration, and User Interaction layers.

The Discovery Pipelines apply unsupervised or self-supervised machine learning to model systematic baselines and detect novel and unusual data. Currently, there are very few AI applications in healthcare that combine both unsupervised methods with pre-trained neural networks. New software packages that extend the common ML frameworks will integrate advanced unsupervised algorithms for pattern discovery.

The Predictor Pipelines perform time-series causal analysis to forecast future conditions, interventions, and medical outcomes.

The Recommender Pipelines will scan through millions of patient treatment records and identify optimally tailored treatment plans (medications and therapies) for individual patients’ conditions. Deep Reinforcement Learning (DRL) recently stunned the world with super-human performance with very high-dimensional state and action spaces. While DRL tools such as PyTorch, tf-Agents, and Azure Cognitive Services already exist, we believe there will be many more DRL tools developed extending common ML Frameworks in the future.

eSimplicity architects our AI implementation in 5 layers. The Data Storage and Query Layer enables the AI pipelines comprehensive access to raw data in the underlying data lake. Before Discovery, Recommender, and Predictor Pipelines can either train a baseline or infer an estimate, the data must be processed through a Feature Extraction Layer. The most valuable information in medical journals, pharmaceutical test results, electronic medical records, patient-generated data, and financial records is common in natural language, and semantic features must be extracted using the latest Natural Language Processing (NLP) algorithms. The Interpretability Layer interacts with the pipelines and the feature Extraction Layers to identify which data features contributed to the prediction or recommendation. While at this time a new framework called Integrated Gradients (IG) can be computed with popular ML frameworks like TensorFlow and Caffe, in the near-term, software packages will be available to easily apply IG which will standardize interpretability within all ML applications. The Model Orchestration Layer contains a set of tools that enable efficient training and deployment of high-performance models. The User Interaction (UI) Layer enables a medical provider to review relevant raw data, predictions, or recommendations within the same contextual view of a dashboard with data visualization.