Sequential Postoperative Surveillance for Anastomotic Leak Using Dynamic State Estimation

• Sompong Rattanakorn

  Department of Computer Engineering and Digital Technology, Buriram Rajabhat University, 439 Jira Road, Nai Mueang Subdistrict, Mueang Buriram, Buriram 31000, Thailand

  Author

• Chaiwat Lertsakul

  Faculty of Information and Communication Technology, Yala Rajabhat University, 133 Tessaban 3 Road, Sateng Subdistrict, Mueang Yala, Yala 95000, Thailand

  Author

Anastomotic leak remains one of the most difficult postoperative complications to recognize early because its emergence is temporally distributed, physiologically heterogeneous, and operationally entangled with routine recovery after major gastrointestinal surgery. Clinical deterioration rarely appears as a single abrupt observation. Instead, it unfolds through a sequence of weakly informative signs whose interpretation depends on timing, co-occurrence, persistence, and response to treatment. This feature of the problem creates a mismatch with computational approaches that estimate risk only before surgery or from a single static postoperative snapshot. A surveillance perspective is more appropriate because it treats leak detection as repeated inference under evolving evidence. This paper develops a technical framework for such surveillance using structured electronic health record data. The proposed formulation models postoperative recovery as a partially observed stochastic system in which latent anastomotic integrity, inflammatory escalation, perfusion disturbance, and support intensity evolve over time while measurements arrive irregularly and with informative missingness. Risk is updated sequentially through dynamic state estimation and horizon-specific event prediction rather than through one-time classification. The paper also derives action-aware alert rules that account for false-alarm burden, timing value, intervention cost, and clinician attention, and it outlines validation strategies centered on lead time, calibration drift, and portability across institutions. The resulting argument is that postoperative surveillance should be treated as a longitudinal decision system in which preoperative information provides an initial prior but not the final answer. The goal is not maximal algorithmic complexity for its own sake, but a disciplined representation of the way postoperative evidence actually accumulates in clinical care.

• PDF