Linear accelerator (linac) downtime analysis assisted with a Large Language Model (LLM)

Linear accelerators (LINACs) are critical components of modern radiation therapy, requiring consistent operational performance to ensure uninterrupted patient care. Unplanned downtime not only disrupts clinical workflows but can significantly impact treatment efficacy. Traditional approaches to LINAC reliability analysis have often focused on specific components rather than comprehensive performance patterns. The advent of artificial intelligence and large language models (LLMs) offers new opportunities for analyzing complex, unstructured maintenance data to extract meaningful insights that can inform maintenance strategies and improve clinical operations.

This study aims to analyze long-term operational performance of LINACs by investigating 10 years of maintenance records from three VarianTrueBeam LINACs. We sought to identify fault patterns, quantify downtime durations, determine the most vulnerable components, and evaluate the impact of the COVID-19 pandemic on maintenance practices.

We analyzed 1584 maintenance reports from three Varian TrueBeam LINACs spanning 4-13 years of operational use. Our analysis workflow consisted of three main stages: (1) Data normalization involved extracting and cleaning reports from multiple versions of service documents into a uniform tabular structure; (2) Report classification utilized an LLM with custom prompt engineering to categorize each report into predefined work type and failure type categories; and (3) Descriptive analysis examined patterns and trends in machine performance, including time series decomposition to identify seasonal trends and detailed analysis of service events requiring external technician involvement.

All LINACs consistently maintained operational downtime within the 5% threshold agreed upon with the vendor, with only one instance approaching this limit in 2021. Collimation systems, control hardware, and power systems accounted for the highest proportion of maintenance cases. Replacement and repair were the most common work categories. No consistent increasing trend in failure frequency with machine age was observed, and seasonal pattern analysis showed weak seasonality. Field service engineer (FSE) visits have increased steadily over time, with replacement tasks most frequently requiring external technician support. Post-2020 maintenance showed increased average work hours in most categories compared to pre-2020, particularly for replacement-related activities.

Our findings provide insights into the operational performance of LINACs over time, with implications for clinical budget management and maintenance scheduling. The data analysis methodology and LLM-based classification techniques developed in this study demonstrate an effective approach to analyzing historical maintenance records that could be applied at other institutions.