Cars can be plugged in at the mechanics for electronic diagnosis, customer issues logged in enterprise support systems receive immediate potential solutions to their issue prior to a customer service representative looking at it, and computers send error reports when an application crashes. In industries across the world automated diagnostics becomes more and more prevalent leveraging continually advancing algorithms that become increasingly intelligent in identifying solutions to known problems. Yet in the health care industry Doctors have out dated and limited access to potential solutions and details from a patient’s case are seldom fully available to be investigated holistically.
There has been significant research on automated diagnosis, but limited practical application and integration of systems. The idea, well represented in Magnus Stensmo’s Ph.D thesis, Adaptive Automated Diagnosis, clearly paints the picture of how powerful and needed such an approach is.
Bias in today’s computer assisted diagnosis
“Enter symptom, disease type, test name or code” requests one physician diagnosis database. As with any human search that begins with keywords chosen by the user, bias inherently influences the results. If a Doctor has an assumed diagnosis, they will immediately begin searching for further evidence that their assumption can be validated. And if it isn’t, then they will have missed other potential diagnoses. Additionally, if the Doctor begins searching by symptoms, while these may be accurate, the order or weight given to any one symptom will give a bias toward related diagnosis when in fact, there may be a symptom not given any credit and thus not included in the search. Regardless of whether you consider today’s databases or the older process of researching in books, the results are always influenced by the bias of the researchers’ initial assumptions.
An Automated Medical Diagnosis System
What is needed instead is an approach that minimizes human bias and considers all relevant and irrelevant data in determining a diagnosis. Computer software does this well. With an automated medical diagnosis system, Doctors could be presented with multiple potential diagnoses based on all of the patient’s current and past details. Such a system could be designed for automated medical diagnosis that is based on probability, utility and decision theory (Read Adaptive Automated Diagnosis). Essentially, the computer software could be fed human observations of symptoms, test results, and any machine data collected such as blood pressure, heart rate, oxygen levels, etc. The software would then compare these observations with a database of potential diseases and external agents (e.g. , viruses, bacteria) to determine the most probable diagnosis.
These results would then be presented back to the doctor along with a probability rating indicating which ones are likely most relevant or accurate. Each diagnosis could also then be presented with additional direction to the doctor to further explore for additional symptoms and/or order an additional test. These additional observations and/or test results would again then be fed into the system where it could reevaluate the probable diagnoses canceling out some while raising the probability of others.
In addition to immediate interactions with the software, Intensive Care Unit’s machine observation data (e.g., oxygen levels, heart monitors) could be constantly fed into the system to allow the software to be looking for patterns that match other known diagnosis that would never be able to be caught by a human as it would take too much time to evaluate the data. Nurse’s notes could also be used in a similar manner.
The most challenging data for the automated medical diagnosis software to interrupt would be results of imaging systems. However, current advances in face recognition technology can and have been applied to reviewing images such as x-rays. This is important as “image interpretation is an error prone task. The number of lawsuits filed against medical imaging professionals that are related to the miss of a diagnosis is close to 70% (Berlin, 1995). The most common errors are perceptual errors that lead to diagnoses misses, representing about 60% of the cases (Renfrew et al., 1992).” – Application on Reinforcement Learning for Diagnosis Based on Medical Image . With a software solution, images could be reviewed against known patterns and then presented to a Doctor for a final review. Similarly, following a codification process, a standard could be defined to notate what one sees in an image that would be relevant to a medical diagnosis, which could then be understood by the software.
The database of potential diagnoses should be an online service that all medical institutions interact with in a real-time basis. This would allow for two key additional benefits. The first would be to connect Doctors in real time with their peers at other facilities with patients experiencing a similar condition. This would allow for immediate collaboration that could lead to a faster treatment. And second, by having the system online, you ensure that every doctor has access to the latest scientific diagnoses.
The system should then be designed to become increasingly more intelligent with time from increased information about each disease and ailment as well as feedback as to the accuracy of its results. Specifically, after each diagnosis, the Doctor would be responsible for submitting feedback as to whether or not the chosen diagnosis and subsequent treatment successfully resolved the patient’s problem. These responses could then go through an automated peer review with the results updating the probability factors to each disease and associated symptoms, test results, etc. Additionally, the software would record the other data learned from each patient in consideration of relevance to future cases. Results could be rated in accuracy based on how they are determined. For example, the results obtained from a biopsy or autopsy may be weighted as being more accurate than simply an observation from the Doctor that the patient recovers.
Today there are a number of point-of-care (POC) testing solutions, (also called near patient or bedside testing), but no large scale application of an automated medical diagnosis system. Today’s applications are for very focused blood tests and are far from providing the capabilities described above. Some examples:
The machine data required for the software has been developed such that it could be fed into such a system. There are many systems available today that would likely only require some standards to be developed for the resulting data. Some examples:
Software engineers everywhere can read this idea and will instantly recognize that today’s technology supports this solution. However, the greatest obstacles would be to gain public support for the requirements of fully electronic medical records and Doctors learning to work together. Progress is being made on changing the culture on both of these topics; unfortunately we have a long ways to go.
This article is dedicated to Grace Allen, a beloved mother, grandmother, great-grandmother, friend, cousin, partner and overall amazing woman. Had this system been in existence today, we would have had some more wonderful years with her.
- Stensmo, M & Sejnowski, T. J (1996) Automated Medical Diagnosis based on Decision Theory and Learning from Cases, World Congress on Neural Networks pp. 1227-1231
- Stensmo, Magnus (1995). Adaptive Automated Diagnosis. PhD thesis, KungligaTekniska Hogskolan (Royal Institute of Technology), Stockholm, Sweden.
- Stelmo Magalhaes Barros Netto, Vanessa Rodrigues Coelho Leite, Aristofanes Correa Silva, Anselmo Cardoso de Paiva and Areolino de Almeida Neto (2008). Reinforcement Learning ISBN 978-3-902613-14-1. Application on Reinforcement Learning for Diagnosis Based on Medical Image.
- Point of Care Testing