Data: Maximizing Value with Advanced Methods

Medicine as a Scientific Discipline

Medicine has, since its inception, been a scientific profession. Observations are translated into hypotheses and those withstanding scientific scrutiny are translated into methods for case definitions (used for diagnosing) and interventions. Each clinical assessment involves data. The complexity of medical data is expotentially greater that any other scientific discipline and the inability to full master the data has created a need for the "art of Medicine" or the use of more intuitive and deductive mental processes to reach conclusions. "Expert opinions" are frequently the source of diagnosis and treatment decisions. Yet, the value of expert opinion is seriously challenged when put to rigorous scientific tests (see Evidence Based Medicine).

Modern computing technology has changed the landscape. Computers can perform complex manipulations beyond the ability of the human mind. As a result, there are now systems available for creating belief systems, knowledge environments and decision support that can add great value to Medicine.

Data Extraction

Maximizing the use of clinical data requires methods for extracting and manipulating the data. Often called "data mining," these techniques leverage the value of the relational databases which store the data. Within a few seconds it is possible to extract data from several tables and present it in humanly understandable formats. Optimizing the functionality of these extractions requires foresight in designing the systems and is more difficult as an afterthought. Therefore, the PTTF recommends that data management be a high priority in rolling out the EMR.

Data extraction will be critical in several areas:
  1. Patient safety. Alerts are produced by data extractions that recognize a problem.
  2. Quality Improvement. Similar to alerts, these data extractions point to potential problems (see below)
  3. Research. Data extractions can provide alerts about candidates for research protocols. The extractions can also provide data for analysis in reseatch protocols.
  4. Administration. Good business decisions are based on data (see below).
  5. Decision support. Clinical data can be utilized to populate decision support believe systems (see below)
David A Stumpf, MD, PhD

Quality Improvement

Contemporary Quality Improvement is quite different from earlier concepts of monitoring frequencies of events. This “count and report” approach was intrinsically retrospective and reactive. The focus now is on active improvement, incorporating evidence-based guidelines into practice, measuring clinically significant outcomes, and assuring timeliness of care provision. The electronic medical record (EMR) is an integral part of this improvement effort. Features include:

  1. Automatic aggregation of data in the course of everyday patient care
  2. Alerts for unusual events in order to assure rapid recognition of potential errors and near misses. For instance, prescription of naloxone would be instantly highlighted so as to investigate whether a narcotic was overused.
  3. Assurance of timeliness of care by automatic time stamping. For instance, the sequence of events for admission of a child with cystic fibrosis would monitor time of arrival to floor/first vital sign, ordering of antibiotics, and delivery of antibiotics to the patient.
For this system to be successful, these functions must be considered at the point of initial design. This can most readily be recognized when considering the implementation of Children’s Best Practices. Up-front integration of a Best Practice guideline would ensure that patients are appropriately flagged for inclusion in the pathway, pathway orders are automatically uploaded for completion/modification when the chart is opened, and real-time performance to expectations is tracked. Retrofitting these functions to an existing system would be far more difficult.

The history of data aggregation at CMH is best exemplified by the initial rollout of Carevue in the critical care units. At that time the vendor did not support the desire to use the clinical data repository for data aggregation and reporting. Carevue includes most of the information required to monitor ICU function, such as central line and ventilator days. These are currently gathered manually from two different sources. It is impossible to track meaningful clinical outcomes, such as selecting patients with a specific severity of respiratory failure and comparing ventilator approaches.

The cost-savings of automated data collection over manual gathering is self-evident. Additionally, there is far less likelihood of selection bias and recall bias. Personnel can be redeployed from data collection to actual quality improvement activities. The patient safety features are enhanced by real-time trend analysis. For example, increased use of antibiotics after a specific surgical procedure may herald an increased incidence of postoperative infection, permitting earlier recognition and intervention to minimize this complication.

Thus, the EMR is essential for superior Quality Improvement activities, including
  1. data aggregation and reporting
  2. integration of Best Practices across the continuum
  3. patient safety alerts
  4. real-time trend analysis
  5. monitoring of timeliness of care provision
  6. just-in-time feedback to clinicians and families
  7. compliance with external reporting requirements
Denise Goodman, MD

Business Decisions Require Data

It is currently very difficult for Clinical Practice Directors to obtain business data to support decisions. Extraction of scheduling data has recently reach a degree of sophistocation that creates some useful data. Some of the most important business decisions require linking financial data to operations; this is now not possible, as discussed in another document, The Problem

Integrating patient and physician demographics, scheduling, insurance, and billing data provides a powerful dataset that permits important analyses and conformitity to business rules:
  1. Insurance coverage can be verified at the time appointments are scheduled and proper authorizations obtained.
  2. Documentation can be delivered to referring physicians and all caregivers
  3. Claims can be submitted more easily with required documentation from the same dataset.
  4. Billing can be based on data in the clinical record, ensuring compliance with reimbursement regulations.
  5. Physicians referring only patients with poor reimbursements may be identified.
  6. Physician productivity can be readily determined by linking scheduling to RVUs.
Decision Support: Extending the Value of Clinical Data

At the present time, most clinical data is simply looked up and viewed. Secondary processing is minimal. Extractions of grouped data are limited. By borrowing from other industries and adapting powerful methods to medical data, we can realize great, added value from the data we now routinely collect. This added value can support safer, more efficient, more insightful and generally better practices. Some examples will help.

Bayesian Networks

MIT's January 2004 Technology Review selected Bayesian systems as one of the "10 emerging techologies that will change the world." Bayesian belief systems can be created from raw data, without prior assumptions. They "learn" from the initial data and then continue to learn from each new data entry. Bayesian systems can also be created by experts, who create the initial assumptions, and then refined by ongoing data. They thus transcend the limitations of evidence based medicine and of expert opinion. A simple example will illustrate their power.

The belief system for a car (see figure), designed to evaluate one that will not start, is illustrated in the diagram. The belief system contain nodes that can be evaluated. Nodes are linked together based on the assumptions of the belief system. A cost can be attributed to testing each node. As a test of the battery, it would be less expensive to turn on the lights rather than replacing the battery. Once constructed, this sytem can ask you questions. If you start with only an understanding that engine will not start, it asks whether the engine turns over. If it does not, then it asks whether the lights turn on. If it does turn over, it will ask for a reading from the gas gauge. With each response it will give the probability for each node as the potential source of the problem. Thus, before you replace the starter, you will have a high probability that it is the problem.

The Bayesian belief system provides some interesting insights. These are often not obvious from the raw data (the battery is essential for starting), but are intuitively clear upon reflection. The battery is not directly linked to the engine starting in the illustration and our conclusions about it can be quite accurate with looking directly at it.

Similar belief systems can be created for a medical issues, such as weakness. Creating such a system changes ones perspective on the problem and focusses attention on the problem solving aspects of the task being pursued. Traditionally physicians look at the spinal cord, anterior motor neuron, nerve and neuromuscular junction as a continuous pathway directly connected to the muscle. In the Bayesian modeling, the deep tension reflexes emerge as an early question. This is intuitive to the experienced clinician, but clearly evident to everyone with this construct.

In this example the Bayesian belief system was populated by estimates of the relative probabilities by an expert. It could be populated by actual data from a set of patient's clinical records. Once these probabilities are available, predictions can be made. However, further information would help prioritize the questioning. Some things are easy or low cost; others are painful or expensive. With a fully informed Bayesian belief system, these factors would be considered and the least expensive, risky or painful course selected. These selections would perhaps suprise a clinician, but they would be intuitive upon reflection.

The next few items illustrate how observations alter probabilities.

The initial probabilities, given only weakness indicate the relative frequency of diagnoses of muscle, nerve, anterior motor neuron (AMN), and spinal cord.
Given the lack of reflexes, the probabilities (in a child) shift strongly to the anterior motor neuron.
A distal sensory loss changes everything and makes the nerve the most likely localization.
An excellent study from Intermountain Healthcare is also instructive. Lagor, et al. trained and tested a Bayesian network and an artififical neural network from data on 32,662 adults with presumptive community acquired pneumonia (CAP). Both networks had excellent accuracy in predicting qualification for the CAP treatment guideline. They utilized Netica software for their research.
Bayesian systems can be used for a wide variety of tasks, including diagnosis, qualification for guidelines or clinical pathways, risk recognition, and administrative management. Deploying these advanced systems will set CMH apart, providing marketplace differentiation and advantage while, at the same time, improving care and the academic opportunities. Do this will require the development of a formal informatics program (see below).

David A Stumpf, MD, PhD

PTTF Recommends the Establishment of an Informatics Program

An Informatics program conducting research into the benefits of new technologies at Children’s Memorial Hospital would create knowledge about the best ways to use these technologies, and would help justify the costs. Support for this program can be obtained from external sources, such as the National Library of Medicine. NLM informatics grants have already been extended to the Children’s Hospital of Philadelphia, and also to University of Chicago. An Informatics Program would enable Children's Memorial Hospital to be competitive in this important research arena. Michael Miller, MD