Wednesday 17th October, 2012
2:05pm to 2:45pm
Trends over the past two decades indicate that the quantity and precision of diagnostic data available for a single patient has increased dramatically, the amount of published medical knowledge is doubling every few years, and a number of promising therapies have been developed. Despite all these advances, medicine remains largely mired in a ‘one size fits all’ paradigm that has led to an explosive increase in patient costs without a concomitant improvement in patient care.
We are on the verge of a paradigm shift in healthcare. Traditionally, medical knowledge has being derived from carefully conducted clinical studies, namely evidence-based-medicine; now, a new form of evidence is emerging – that created by rapid learning systems that will mine vast amounts of electronic patient data collected in routine care, to create “evidence generated medicine.” Thus, mining the millions of patient records collected routinely in the daily care of patients has tremendous potential to individualize care to the specific patient.
In this presentation, I will describe a first-of-its-kind US/Euro health IT network consisting of 10 cancer centers in 5 nations. In this network, cancer centers are able to securely learn personalized models from patient data collected across all centers. Learned models for predicting patient survival and side effects for 3 different cancers (lung, rectal, larynx) have been made available to the public and physicians at www.predictcancer.org.
Creating models from patient data collected across multiple centers provides statistical power, but leads to several challenges:
1. The foremost among these is to protect patient privacy. We have developed privacy preserving data mining methods that allow us to securely mine all patients in the network, while ensuring that all patient data remains with the firewalls of each institution. We are able to derive models that have better performance than models based on smaller data sets from individual institutions.
2. Patient data that has been collected for routine care have many problems from the mining perspective. The data are noisy, have errors and omissions, and in many cases, are biased due to the variance in care across nations and between and within institutions. We deal with these issues by leveraging the large numbers of patients in our multi-national population.
3. Mining patient data is a multi-data-source problem. Rich clinical data is available not just from structured demographic, lab and drug databases, it needs to be extracted from unstructured sources, such as medical images, treatment plans and various omics. All this information needs to be leveraged to learn better, more personalized models.
4. Each institution stores patient data in a variety of multi-vendor source system and in different formats. Each institution collects different kinds of patient data, at varying levels of detail, in different languages and uses varying terminology. Our approach allows us to normalize data across centers on an as-needed basis, thus reducing the normally-intractable problem of mapping all data to a manageable one.
5. Finally patients are unique. Our system has to scale to learn from patients with very little data (e.g., on their first visit) as well as from patients who have records spanning decades.
In this presentation, we will describe how we overcome these issues to learn personalized models that have been statistically validated and published in leading conferences and journals. Additionally, we describe how pharma companies can mine these patient records to more efficiently find patients for clinical trials. The majority of the talk will present case studies and results that illustrate some of the challenges and opportunities unique to mining healthcare data. We conclude with a glimpse of a more-efficient healthcare future, where treatment decisions are driven by evolving knowledge that is continuously mined from patient records collected in health systems all over the world.
Head, Center of Innovations, Siemens Health Services
Sign in to add slides, notes or videos to this session