Case-Based Reasoning in the Health Sciences: Foundations and Research Directions - Computational Intelligence in Healthcare: Advanced Methodologies - page 124

efforts in the field focused on modeling medical expertise, especially for diagnosis,
treatment planning, and follow-up care. This was in line with the early AI goals
of representing and utelizing the knowledge of human experts.
The first proof of concept that CBR could be used for health sciences appli-
cations is attributed to Janet Kolodner, who wrote the CBR textbook [11], and
Robert Kolodner, a psychiatrist [19]. The two met by chance at a conference,
where discovering that they shared the same last name, they began discussing
their interests in computing and psychiatry. In [19], they proposed SHRINK, a
case-based system to aid with psychiatric diagnosis and treatment. They aimed
to establish a framework for clinical decision support systems that could also as-
sist in the teaching of clinical reasoning. They presented actual psychiatric cases
to demonstrate how these could be used to reason and to build expertise. While
SHRINK was never fully implemented, many of its ideas were later implemented
in MNAOMIA [20], a CBR system for diagnosing and treating psychiatric eating
disorders.
MNAOMIA implemented the Retrieve, Reuse, Revise, Retain cycle shown in
Figure 1, but added extensions to deal with challenges from the psychiatric do-
main. These extensions enabled the system to: reason from both episodal, or
case, knowledge, and theoretical, or documentary, knowledge; assist with clin-
ical tasks including diagnosis, treatment planning, follow-up care, and clinical
research; use cases acquired by the system to inform clinical research; and reason
over time to support patient follow-up. A psychiatrist involved in the MNAO-
MIA project shared his vision of how a computer system could best support
his clinical practice. He wanted an electronic repository for his patient records,
storing details of medical history, environment, symptoms, diseases, treatments,
tests, and procedures. He wanted online access to patient records of the other
psychiatrists in his practice, as well. He envisioned being able to consult the
system when he encountered a new patient or a new disease episode. The sys-
tem would retrieve and display the most similar old patients or disease episodes,
note similarities and differences to the current case, describe past successful
treatments, and recommend diagnoses and treatments for the current situation.
Completely realizing this vision of clinical decision support is still an important
goal for CBR in the Health Sciences.
Many of the first CBR systems in health sciences domains were built for
medical diagnosis. Two of the earliest and best-known diagnostic systems were
PROTOS [21,22] and CASEY [23]. PROTOS diagnosed audiological, or hear-
ing, disorders. Its case base contained approximately 200 cases for actual patients
who had been diagnosed by an expert audiologist. A patient's disorder could be
described by up to 58 distinct features, although, on average, values were known
for only 11 features per patient. A patient's diagnosis was a categorical label,
telling which of 15 pre-specified audiological disorders the patient had. Diagnos-
ing a new patient, then, meant classifying the new patient into one of the 15
categories. PROTOS used reminding links to determine the most likely classifi-
cation for a patient. If it found a past patient within that category that closely
matched the new patient, it would extend the past patient's diagnostic label to