John Monson

CSCI 222

Paper Presentation

PIP

Towards the Simulation of Clinical Cognition

Taking a Present Illness by Computer

Who wrote it:

· Stephen G. Pauker

· G. Anthony Gorry

· Jerome P. Kassirer

· William B. Schwartz

What Institution:

· Massachusetts Institute of Technology

Sources:

· June 1976 The American Journal of Medicine Volume 60

· Dr. Keuneke’s notes: www.ecst.csuchico.edu/~amk/foo/csci222/notes/pipabel.html

· Jerry Loyde’s Presentation: http://ai-web.ecst.csuchico.edu/modules/csci222/presentation/loyd/pres/index.htm

What is it:

· PIP stands for Present Illness Program.

· PIP is a framed based diagnostic system.

Task:

· Ask a patient questions and test the patient to diagnose the cause of the patient’s present illness.

Domain:

· Medicine as it applies to edema.

· It consists of 20 diseases and 70 frames.

I believe that PIP is domain dependent because it is specific in certain diseases and is used for only the clinical problems. However, as Jerry mentioned in his presentation, the technology used by PIP (frames, links, daemons) could be applicable to different domains. The authors did point out that PIP’s scope could be broadened, but it would take considerable effort.

Knowledge Representation and Control Structure:

Architecture:

There are four basic components of PIP that interact with each other: Patient Specific Data, Supervisory Program, Short-Term Memory, and Long-Term Memory.

Patient Specific Data

These are the facts provided by the user either spontaneously or in response to questions asked by the program.

The Supervisory Program

The supervisory program guides the computer in taking the present illness and oversees the operation of various sub processes, such as selecting questions, seeking and applying relevant advice, and processing algorithms.

The supervisory program generates and tests hypotheses and employ information gathering activities.

Short-Term Memory

Short-term memory is where patient data interacts with general medical knowledge stored in long-term memory.

The general medical knowledge that is moved into short-term memory that interacts with patient specific data is determined by the supervisory program.

Long Term Memory

Long term memory contains knowledge about the diseases called frames.

Frames:

Frames are a collection of knowledge about a disease stored in Long Term memory. Frames contain the information necessary for PIP to form hypotheses. There are three types of frames:

· Disease Frames

· Clinical State Frames

· Physiological State Frames

Frames contain the following:

· Links to Other Frames

· Typical Findings

· Rules to Rate Hypothesis score of criteria being met

· Rules to suggest other hypotheses for differential diagnosis

After the complaint has been characterized and all relevant advice has been acted upon, the supervisory program proceeds to generate working hypotheses. This includes moving frames from long-term memory into short-term memory for further evaluation and comparison with the patient’s complaint.

Frames can exist in one of four states:

· Dormant

· Semi-Active

· Active

· Accepted

Daemons

Daemons are small programs that are associated with findings in frames. Daemons try to match facts and may move frames from long-term memory into short-term memory.

Control

Question Asking

The program begins by asking the user for a “Main Complaint”. Based on the Main Complaint, the Supervisory Program is able to decide what further questions to ask.

Hypothesis Generation

PIP searches the disease frames in LTM to find matches for the user’s complaint. It then drags frames that match into STM.

PIP needs to limit the number of hypothesis under consideration. PIP controls this by:

· Only searching frames that are one link away from the frame in STM.

· Frames with a low score are removed from STM and frames that are semi active are not allowed to run their daemons.

Hypothesis Testing

Hypotheses generated by the program are evaluated to determine the extent to which they constitute reasonable explanations for the patient’s condition. This involves two steps:

· Determining if the hypothesis fits the case.

· Examining each hypothesis to determine the extent to which it can account for all of the facts in the case.

PIP uses a scoring algorithm to rank the hypotheses and find matches or reject a hypothesis. There are three cases:

· Exact Match

Every important feature of the prototype is found in the case and there are no features of the case which are not explained by the prototype.

· Case too Big

There are features of the case which are not explained by the hypothesis.

· Hypothesis too Big

There are features in the hypothesis that are not found in the case.

The process of generating hypotheses, testing them, and moving frames into and out of STM is repeated until PIP reaches a conclusion regarding the patient’s symptom.

Sample Run:

Jerry Loyde provides a beautiful, interactive flow chart detailing a sample PIP session at:

http://ai-web.ecst.csuchico.edu/modules/csci222/presentation/loyd/pres/control.htm

PIP And Other Expert Systems:

Like Mycin on MDX, PIP is applicable to the study of medicine. But that is where the similarities end. The internal structure of the expert systems are different. Mycin is used for medical diagnosis. Mycin is a rule-based backward chainer. MDX, on the other hand, is a hierarchical specialist classification system. MDX is a static system, where PIP is dynamic. Pip is a goal driven, forward chaining expert system. Pip uses a breath first search, because there is an algorithm that decides which it should do.

Limitations:

Jerry Loyde notes several of shortcoming pointed out by Dr. William B. Schwartz in a later article about PIP. His comments can be found at the following web address:

http://ai-web.ecst.csuchico.edu/modules/csci222/presentation/loyd/pres/limitations.htm

· Because PIP was developed in the 1970’s, I assume it was very limited by the technology at the time. For example, I am sure the storage devices used for LTM were only able to hold small amounts of data(8K) in comparison to todays standards. This would severely limit the amount of disease frames that could be held in LTM and would limit PIP in its ability to diagnose a wide range of diseases. The article on PIP mentioned that the cost of storing the knowledge base could cost as much as $20,000. However, the authors believed that advances in technology will be able to facilitate storing two million facts.

· Processor speed would also be a limiting factor. I can only imagine how slow the process of hypothesis testing and hypothesis generations was with the slow processors used in the seventies.

· PIP is currently only applicable to the problem of edema. It would take considerable effort to adapt PIP to a wide range of clinical problems.

· It is difficult problem to organize, retrieve and apply the relevant data of frames.

Goodness:

· Although it would take much effort to increase PIP’s scope of diagnose, it could be applied to a larger range of clinical problems. Many strategies other than those currently employed would have to be uncovered if the system were to deal effectively with other clinical problems. This would involve numerous medical and real-world facts to be added to the program and new techniques to deal with multiple coexisting diseases.

· The frames in LTM are organized in a very smart fashion. They are organized into a network which facilitates efficient retrieval of closely related blocks of information. This undoubtedly helped to compensate for the lack of processing power that was available for PIP in the 70’s