An Integrated Supervised and Unsupervised Machine Learning Framework for Classification and Clinical Pattern Discovery in Musculoskeletal Outpatient Records in Libya

Abstract

Abstract

This study aims to employ supervised
and unsupervised machine learning to analyze
outpatient
clinic
records
for
musculoskeletal
disorders in Libya. The study relied on real-world
clinical data containing structured fields and free
clinical text. Three research datasets were derived
from the original data: a supervised dataset for
classification, an unsupervised pre-diagnosis dataset
(PRE), and an unsupervised post-visit dataset
(POST). The unsupervised learning results showed
that the PRE dataset revealed three broad clinical
groups, whereas the POST dataset revealed five
more detailed groups. In the supervised learning
pathway,
LinearSVC
achieved
the
best
performance, with an accuracy of 71.96% and a
Macro-F1 score of 71.87%. Accuracy further
increased to 84.16% when only high-confidence
cases were classified. The findings indicate that
unsupervised learning is useful for understanding
clinical structure, whereas supervised learning
provides
practical
classification
capability.
However, overlap between clinical categories and
low-information clinical texts remain among the
main performance challenges.
Keywords:
electronic
health
records,
musculoskeletal disorders, supervised learning,
unsupervised learning, clustering, Libya