Menu Close

Dr Paul Wallace Interview

By Dr Tom Foley, Dr Fergus Fairmichael.


Paul Wallace, MD, is Chief Medical Officer and Senior Vice President for Clinical Translation at Optum Labs. Before joining Optum Labs, Dr. Wallace was senior vice president and Director of the Center for Comparative Effectiveness Research (CER) at the Washington DC based Lewin Group.  Dr. Wallace is board vice-chair for AcademyHealth and a board member of the eHealth Initiative. He is a former Medical Director and Clinician with Kaiser Permanente.

About Optum Labs

Optum Labs was founded as a partnership between Optum and Mayo Clinic in January 2013, It brings together a community of over 20 partners with the aim of improving patient care by sharing information assets, technologies, knowledge tools and scientific expertise.  Optum Labs is in the unique position of having access to a patient level de-identified clinical database from the electronic health records of around 40 million people as well as an additional 150 million patients from claims data, of which there is an overlap of around 20%.  Optum Labs has been able to link data together at a personal level while preserving the de-identification of this database.


The learning health system

The learning health system is sociotechnical. It has technical underpinnings and its engineering is an interesting exercise, but the major change required for its development is social and organisational.

The LHS concept of research innovation, turning data into knowledge, being good stewards of data and translation into practice, will take time to develop. There is a need to provide clear use cases and proof of concept projects, delivering good science, that gradually grow the user community. This will, in turn, lead to bigger impact use cases.

Changes to research

Over the last five or ten years, research has reached an inflection point. In 20th century medicine, a great deal of the cost of clinical trials was associated with the data collection. 

“Research is changing from a hunter/gatherer mode, where huge amounts of effort is invested to associate data with rare events, to a harvest mode in which huge amounts of data are used more efficiently to give insight.”

There is great reliance on the clinical trial because it is seen as the most risk averse approach. It reduces concerns about bias, but sometimes this distracts from the relevance of the original question and has meant that we have not been focussing on what is important to patients.  The LHS is allowing us to develop a greater understanding of patient data.  Each patient’s data consists of their admin claims, clinical records, their socio-economic status, personal values etc. This better understanding of patient data can account for a lot of the variability we have tried to control in clinical trials. 

Research is currently very inefficient.  There are many drawbacks to randomised control trials.  RCTs can be great if the patient in front of you meets the criteria for the trial, but real word patients often don’t fit into the same boxes as those from RCTs.   This does not necessarily mean that RCTs are wrong, just that they are a limited use case and may not be applicable for more complex cases.  There is a need to understand the cases when clinical trials are still needed. 

Observational studies are cheaper, we can perform hundreds of studies for the price of one RCT.  The question becomes “How well can we replicate trials in an observational environment?”  It could also be viewed as a moral imperative. If you can do a trial in silico vs in vivo, then there is no risk to the participants.  

Increasingly, each patient is a rare event due to their combination of comorbidities and demography.   This leads to the question becoming “how can we find many N of 1’s like our patient?” and moves away from the idea of applying the population mean to our patient, which probably has very little relevance to them.  This can be addressed now that there is an abundance of data and we begin to move to a harvesting mode.  This approach is a step towards personalisation. 

Data sources

Healthcare and education stand out as two industries that have not updated how they use electronic data. Other industries have recognised that the use data can be a powerful driver for improvement in services. For example, amazon has made great use of your shopping habits to suggest what you may want to buy next. 

Claims data is often the backbone to current systems and provides the most comprehensive picture of patient interactions with the health system.  This can lack detail, which would ideally come from EHRs, but some data may still be missing. For example, when patients are seen by more than one provider.  In these circumstances, attempts can be made to impute the missing data and it can be decided whether it is sufficient to conduct the necessary research.

There are differing approaches to obtaining the data:
1. On the input side you can use highly structured inputs that are likely to have high data quality but also likely to have gaps. 
2. Another approach would be the use of Natural Language Processing (NLP) that can help to fill these gaps. NLP is improving and is very good even now.  Concerns over lost context and the syntax of how the elements fit together are also improving.

Optum Labs already use both of these approaches.  Around 70% of data used in Optum Labs is codified using NLP.  The experience has been that clinicians do not want structured fields.


There is not yet a perfect solution to the challenge of interoperability.  There are many different approaches such as producing standardised coding systems and data structures.    Interoperability for research and interoperability for patient management are very different things with different needs and purposes.  In interoperability for research, data de-identification, privacy and confidentiality will be important. It will likely need different sets of approaches depending on the use cases. 

Translation into practice

“Translating the findings of data analysis and research into a change in practice can be a real challenge and the level of effort depends on what you are doing.  If it is a simple process change it can be done quickly and efficiently, for example a formulary change has a level of effort of one.  Introducing a new concept that begins to change behaviour, such as a guideline, increases that level of effort to about ten.  Changes that go against peoples beliefs move up to an effort level of one hundred.  Asking a practitioner to change their own behaviour in a fundamentally different way, for example supporting shared decision making, may have a level of effort even higher than this.”    It is important that the study design should have the end user in mind to help minimise this level of effort.  The most successful approaches are those that aim to complement and understand what the clinician already does.  The end user should be involved early and should guide the process.  Ultimately the end user will be the patient and not the doctor, and if you are not designing systems with the patient in mind then it will probably fail.

Risk models, which are used to predict the risk of developing outcomes associated with disease, have traditionally been used to help guide practice, however there is a need to go beyond this and move into decision support. The LHS could effectively give a clinician access to the experience of thousands of other practitioners, like grand rounds but on a larger scale. At the same time, it will be more successful if it preserves the physician’s role in forming a relationship with the patient and sharing the decision making.

Workforce implications

In the development of modern medicine the role of clinician has become disassociated from the role of researcher. Researchers are there to ask questions and supply knowledge which subsequently trickles down to the clinicians.  This approach is no longer sufficient to meet the needs of the front line.  The development of a learning health system means that the skills mix in the healthcare system will have to change. Front line clinicians will become more involved in generating knowledge and the distinction between researcher and clinician will become blurred. This will facilitate a cycle of learning.

IT should complement and extend the capabilities of the healthcare professional. People do not want to see “Dr Watson”, they want to see a doctor who is backed up by the knowledge of Watson. It is important not to forget the unique qualities a doctor brings to the relationship with the patient that cannot be offloaded to the IT side, such as empathy, intuition, autonomy as a doctor, and accountability to the patient.


A learning health system is achievable. The concept is here now, and there is no reason why it could not be in place in 10 years, but it would be extraordinarily disruptive.  The technology already exists for the rapid querying of clinical systems and it is already possible to incorporate decision support tools into EHRs. Computing ability is not the limiting factor. Creating the culture of change is likely to be much harder.  The learning healthcare system will only work if everyone is a learner.  There is a need to recognise how much knowledge is thrown away by not taking advantage of data.