How Does Lumiata Work?

Lumiata predicts the evolving risk of major disease conditions over time for each individual within a population. For each individual, a “Risk Matrix” of future associated conditions or “Predictions” is created, when their health data is applied to Lumiata’s predictive models. Additionally, for each Prediction, a corresponding “Clinical Rationale” representing — key supporting data points is generated. The Clinical Rationale represents the transparent evidence supporting the risk each predicted clinical condition. A model probability for each Prediction is also included to support ranking and sorting of Predictions within the population.

The following illustrates how Lumiata’s AI converts raw data into insights:

Making Sense of Health Data

A Medical Graph providing global clinical context

Personalized medicine requires the ability to interpret relationships among various factors: medical, behavioral, social, environmental, demographic, etc. Where does the knowledge required to interpret data points exist? Unfortunately many locations. For example:

  • Current and past medical knowledge is recorded in medical textbooks and in the minds of physicians.
  • New research and findings are published daily in scientific and medical journals.
  • Medical guidelines are generated regularly by physician groups, medical, colleges and industry experts.
  • The definitions and hierarchies of medical knowledge are represented as public ontologies.
  • New sources of data are emerging daily, from devices and genetic screens to new forms of diagnostic labs.

All these data sets are useful on their own, but when ingested together, and compared to tens of millions of medical records, the potential opportunity for substantially greater personalized medicine emerges.

Lumiata has collected and combined the silos of medical knowledge into a single data structure – the Lumiata Medical Graph. It models medicine and clinical science as a data structure and inference engine. The graph provides inferential clinical context when patient data is queried against it. Currently, the medical graph is composed of hundreds of thousands of unique medical concepts from medical guidelines & protocols, scientific literature, models of disease states, diagnoses, symptoms, and therapies.

Lumiata continually ingests new forms of medical data for inclusion within the Medical Graph. To date, the Lumiata Medical Graph combines and describes relationships among:


Longitudinal patient medical records


Unique medical concepts


Unique medical codes across 10 ontologies


Unique relationships among these concepts


Hours of physician derived knowledge


Medical publications

Understanding the Clinical Context of an Individual

A Simplified View of the Medical Graph for a Single Patient

Imagine a male in his late 30’s who is at a risk of developing Diabetes Mellitus type II (DM2). One way of visualizing this individual’s risk would be to draw the following:

From population statistics, we know that a Caucasian male in his 30’s has about a 7% chance of contracting DM2 in his lifetime. While that figure represents the risk among the entire population of Caucasian men, it tells us very little about this specific individual. Following is an expanded version of the aforementioned example:

We’ve added some new data about this individual - his BMI & his family history of DM2. With this new data we might find that his chances of developing DM2 in his lifetime goes up significantly higher than previously assumed. We have taken the first step from population health to personalized health. But this only scratches the surface of what we might add to further understand the individual. There are many, many other factors that might influence a condition like DM2 (or any other disease). To represent all of these factors, we have developed the world’s largest medical knowledge graph.

Interpreting the Future Part I

From Graph to Model to “Risk Matrix”

To transform the collection of medical knowledge in the graph into actionable insights, Lumiata has developed the world’s first Medical AI engine, which extracts relationships from the graph that become feature parameters of the model. At the same time, our models can find patterns and features from data that are not known in the literature. By leveraging medical AI with the sum of human knowledge about medicine, we can determine an individual’s risk of being diagnosed with a specific condition within various timeframes.

To train our predictive models for an individual’s specific diagnosis risk, we take a subset of longitudinal individual lives with and without diagnoses. Individual lives are a collection of medical records (in FHIR), as well as augmented features, measures, data points, or other analytic outputs from the Graph or other data sources. To build the one-year model for a specific diagnosis, Lumiata removes all the patient’s data from the diagnosis in the previous year. For individuals without a claim for the diagnosis, we remove the end of the medical record from the previous year. That data is then split into test, training, and validation sets.

The data from the training set is used to teach the Lumiata AI engine how to determine the risk probability of future diagnosis. Our models leverage advanced deep learning techniques to generate highly accurate predictions. The precision (PPV), recall (sensitivity) and false positive rate of our models is validated on the training set. For every condition we are interested in predicting, we generate a series of models that can predict a diagnosis now, or up to two years in the future. The result is a Prediction for each associated condition.

Interpreting the Future Part II

The Clinical Rationale Converts Risk into Actionable Opportunities

Knowing a patient is at risk for a chronic condition is only half the story. To provide outstanding care we also need to know why a patient is at risk. The Graph makes this possible by providing a Clinical Rationale behind every prediction.

We leverage the relationships in the Medical Graph to expose all the factors that influence a patient’s risk. This Clinical Rationale is personalized for a specific patient and a specific diagnosis at a specific point in time. As such, it is not just the foundation for understanding a prediction, but it also provides the basis for action.

The Clinical Rational contains the relevant clinical evidence which supports the predictions — includes both causal and correlative evidence. The Clinical Rationale is represented by four categories extracted from the individual’s health data:

  • Relevant Conditions
  • Procedures
  • Abnormal Results
  • Relevant Medication

Driving Value from Data in Real-time

The FHIR API Enables Interconnectivity

The data that Lumiata ingests goes through the Clinical ETL and is then normalized into FHIR. From there, the Risk Matrix and rationale are computed. The output is the full patient representation along with risk (and other analytics), normalized into a single, searchable record that can easily be incorporated into other applications or workflows, or used as a standalone application. With Lumiata's clinical ETL and FHIR API, all of your data is standardized and in one place, together with high quality risk predictions and supporting clinical rationale.

The FHIR API also has population search capabilities that easily query a variety of searches. For example:

  • Which patients are at risk for CKD in a certain location, or under the care of a certain provider or group of providers?
  • What is the relative risk of a population of patients who are under the care of a specific provider?

How Lumiata Standardizes & Normalizes Data

Real-time clinical ETL (“Extract Transform Load”) of medical data

To make the most accurate predictions, we need to know as much about a patient as possible, which in practice means as much normalized data as possible.

The Lumiata Clinical ETL transforms data from a wide variety of sources into FHIR (Fast Healthcare Interoperability Resources), the emerging standard for medical data representation, exchange and interoperability. The FHIR open standard has widespread application and support in the industry. Because our ETL can transform other sources into FHIR, and our Medical Graph and models operate on FHIR, our Medical AI engine is able to process and evaluate most current and future data types, ensuring the best possible predictions now and in the future.

Our streaming technology can handle near-real time data ingestion. This continuous ingestion ensures that as new data comes in, the risk calculations are updated almost immediately and the end user receives the most up-to- date view of the individual.

Seamless Integration of AI-Powered Analytics

Lumiata's core technology brings all of your data together into one standardized format, enabling you to use the data to understand the future risk of each individual and why that individual is at risk. It also provides insights into the current and future risk status of your overall patient population. Lastly and perhaps more importantly, it makes the data and insights easy to access and integrate into your workflow.

Unified. Interpreted. Convenient.