Surgical Process Discovery for Scheduling: Structural and Temporal Analysis of Perioperative Workflow

Date of Submission

Summer 2026

Supervisor

Dr. Tariq Mahmood

Committee Member 1

Dr. Naveed ur Rehman, AKUH

Committee Member 2

Dr. Sohail Imran, KIET

Degree

Master of Science in Data Science

Department

Department of Computer Science

Faculty/ School

School of Mathematics and Computer Science (SMCS)

Keywords

Process Mining, Heuristic Miner, Bayesian Inference, Surgical Scheduling, Perioperative Workflow

Abstract

Surgical departments are among the most operationally complex units in any hospital, where coordination across operating rooms, surgeons, and clinical staff directly shapes patient outcomes and institutional efficiency. Surgical scheduling at most tertiary hospitals relies on fixed, procedure-based time estimates, despite substantial, undocumented variability in case duration across surgeons and specialties. Hospital information systems already record timestamped perioperative events for every case, yet this data is rarely analyzed to establish a reliable process baseline, separate genuine rare pathways from noise, or quantify where delays concentrate.

This thesis applies process mining, principally the Heuristic Miner, to reconstruct real surgical workflow directly from event log data. Unlike prior work, which treats structural discovery and duration analysis as separate exercises, this thesis connects the two: discovered pathways are annotated with performance data so that structural reliability and timing delay are examined together. Baseline discovery produced an interpretable but sparse model with no way to assess the reliability of infrequent transitions. This motivated a Bayesian extension to the dependency measure, a Beta-Binomial posterior mean embedded within discovery itself, validated on a real-world dataset of 162,286 perioperative cases. The Bayesian model surfaced nearly double the edges of the classical baseline, distinguishing a confirmed core pathway from high-confidence clinical exceptions and low-frequency edges flagged for review, including a checklist-sequencing deviation invisible to the frequency-only baseline. Temporal analysis further revealed a severe mean-median divergence on a key transition, independently corroborated elsewhere in the literature, with delay patterns differing sharply by specialty.

These findings were translated into operationally realistic recommendations spanning structural, temporal, and governance considerations, showing that a statistically calibrated extension to an established discovery algorithm can surface process insight that frequency counts alone cannot.

Document Type

Restricted Access

Submission Type

Thesis

Link to Full Text

Share

COinS