PIP

PIP

PIP (1976) is a frame-based diagnostic system developed by Pauker, Szolovits, et al at MIT.

Knowledge Representation

Hypotheses stored in "LTM" are frames of one of three types:

• disease frames
• clinical state frames (common to > 1 disease)
• physiological state frames (common to > 1 clinical state)

• links to other frames
  --> forming a network of frames
• typical findings
  --> used to ask pertinent questions
• rules to rate hypothesis score of criteria being met
• sufficiency and exclusionary rules
  --> rules that automatically establish or reject hypothesis
• rules to suggest other hypotheses for differential diagnosis

• findings that are sufficient to conclude the disease; i.e., pathognomonically TRUE findings
• findings that are not associated with the disease (this is a list of pathognomonic NEGATIVE findings)
• findings associated with the disease
• differential diagnosis slot; a slot to activate other diseases that can be confused with this disease (this is interesting - it really serves as a check to make sure that two diseases that have somewhat similar presentations are BOTH examined if one is brought up)
• links to other diseases; if this disease is established, then any other diseases 'linked' via this slot gain some additional weight
• triggers: that is, findings that suggest the disease
• scoring algorithm: a function to map the findings to a confidence for the disease under scrutiny

Control

PIP uses a "hypothesis driven" control strategy for diagnosis. At the beginning of a PIP session, a list of all presenting complaints are given to PIP. The system then examines all its disease frames and if any initial complaint matches a trigger for a disease, that disease is activated; i.e., hypothesized

Next, the must-not-have slots are examined, and if any active disease has a must-not-have slot value matching a finding, that disease is deactivated.

Finally, all the active hypotheses are scored according to their resident scoring algorithm.

The above describes what might be called one machine cycle on a PIP virtual machine.

At the end of each "machine cycle", PIP can establish a disease in one of two ways:

First, it can use a numerical threshold based on the results of running the scoring algorithm

Second, it can use the categorical knowledge stored in the findings-that-are-sufficient-to-conclude-this disease slot

At the beginning of each machine cycle, PIP will ask questions about the current leading hypothesis. Its method of asking new questions is very simple. THe findings slot of the diseases are ordered by whoever makes up the system to begin with. PIP simply goes to the first finding that is unmarked in the current leading disease hypothesis and asks a question that will fill in that finding. Following that, the machine cycle , as described above, will be reexecuted

Comments on PIP

PIP seems to be on a middle ground between Mycin and MDX. By representing each disease as a frame, PIP can include more information about how to evoke and establish a given disease locally to the disease itself than can be done in a Mycin-like system. In fact, the information stored for a PIP disease frame is not unlike the knowledge that is represented in an MDX conceptual specialist. In particular, the locally defined scoring algorithm in PIP could easily be cast as the typical table structures found in the MDX disease nodes.

But, PIP frames are not viewed (or used) as part of any higher-level structure. Hence, in PIP, there is a fall back to a global mechanism at the end of each PIP cycle that will determine what to consider next. There is little or no notion in PIP of "pursuing a line of inquiry" as there is in MDX (more like activation sets in semantic nets)

Question asking in PIP is controlled by compiling what questions to ask about each disease when that disease is the leading contender among the active hypotheses. Note that although very simple, this method of test ordering is totally independant of most context that is set up by the operation of PIP earlier in a current session. Also, note that the question asking could appear to be quite unfocused in the PIP system as from one cycle to mthe next, new (and potentially very divergent) leading hypotheses pop up.