Research on the Application of AI Agent Technology in Quality Defect Root Cause Analysis of Central Sterile Supply Department

This study demonstrates that AI agent-assisted root cause analysis significantly outperforms traditional manual methods in accuracy, depth, and efficiency when addressing quality defects in Central Sterile Supply Departments, offering a valuable tool for enhancing nursing management and decision-making.

The Gist

The Big Picture: A New "Co-Pilot" for Hospital Sterilization

Imagine a hospital's Central Sterile Supply Department (CSSD) as a high-stakes kitchen. Instead of food, they prepare surgical instruments. Every tool must be perfectly clean, packaged, and sterilized before it goes into a patient. If a tool has a tiny bit of dirt, a bad seal, or the wrong temperature, it's a "defect" that could be dangerous.

Usually, when a defect happens, a team of experienced senior staff has to sit down and play detective. They use tools like "Fishbone diagrams" (drawing a fish skeleton to find causes) and ask "Why?" five times to find the root problem. This paper asks: What if we gave these detectives a super-smart AI assistant?

The researchers tested an AI Agent (a smart computer program that can think and reason) against traditional human-only analysis to see who could solve these "kitchen disasters" faster and better.

---

The Experiment: The "Human vs. Robot" Detective Race

The researchers took 30 real-life stories of sterilization mistakes from a major hospital. They split them into two groups:

1. The Traditional Team: Experienced human nurses and managers used their brains and standard tools to solve the cases.
2. The AI Team: A human operator fed the same stories into the AI Agent, which used its massive database of rules and past cases to solve them.

To decide who won, they brought in a panel of 5 top-tier experts (the "Gold Standard" judges) to grade the answers.

The Results: Why the AI Agent Won the Race

Here is how the AI performed, explained through simple metaphors:

1. Accuracy: The "Sharp-Eyed Detective"

• The Claim: The AI found the real cause of the problem 85.6% of the time, while humans only got it right 72.3% of the time.
• The Analogy: Imagine looking for a needle in a haystack. The humans found the needle most of the time, but sometimes they got distracted by a piece of straw. The AI, however, had a "metal detector" that ignored the straw and went straight for the needle every single time.

2. Depth: The "Deep Diver" vs. The "Surface Swimmer"

• The Claim: The AI dug deeper into the problem. It didn't just say, "The worker was tired." It found systemic issues like, "The training schedule is flawed."
• The Analogy: When a car breaks down, a human might say, "The tire is flat." The AI is like a mechanic who says, "The tire is flat because the wheel alignment was off, which happened because the factory's quality control policy changed last month." The AI went from the surface to the foundation.

3. Speed: The "Express Lane"

• The Claim: The AI solved a case in 18.5 minutes. Humans took 35.2 minutes.
• The Analogy: The humans were walking through a library, reading books one by one to find the answer. The AI was like a super-fast librarian who instantly pulled the exact book off the shelf, opened it to the right page, and read the answer aloud.

4. Creativity: The "Idea Generator"

• The Claim: The AI suggested 1.8 new, unique solutions per case. Humans only suggested 0.7.
• The Analogy: If you ask a human to fix a leaky roof, they might say, "Put a bucket under it." The AI, having read thousands of other roof stories, might say, "Let's install a new gutter system and change the roof material to prevent future leaks." It brought fresh ideas from other scenarios.

5. The "Feasibility" Check: The "Real-World Test"

• The Claim: Both groups were equally good at suggesting solutions that were actually possible to do in the real world.
• The Analogy: The AI didn't suggest "magic." It suggested things that could actually be built or bought. It wasn't just a dreamer; it was a practical engineer.

6. The "New Hire" Effect: The "Training Wheels"

• The Claim: When a beginner (a novice) used the AI, their results were 78% as good as the experts. Without AI, a beginner was only 62% as good.
• The Analogy: Think of the AI as a set of high-tech training wheels. A new rider (novice) might wobble a bit, but with the AI's balance, they can ride almost as smoothly as a pro. Without the AI, the new rider falls over much more often.

What This Means for the "Kitchen" (The Hospital)

The paper concludes that the AI Agent is a powerful tool, but it's not there to replace the chefs (the nurses). Instead, it acts as a super-assistant.

• It reduces stress: The humans felt much less mental pressure (cognitive load) when using the AI. It was like having a co-pilot take over the navigation while you focus on flying the plane.
• It saves money in the long run: While buying the AI costs money upfront, it saves time and training costs later. It's like buying a high-end coffee machine: expensive at first, but cheaper than buying coffee every day for a year.
• It standardizes quality: No matter who is working the shift, the AI ensures the analysis follows the same perfect steps every time, preventing human errors caused by fatigue or forgetfulness.

The Bottom Line

This study shows that in the high-pressure world of hospital sterilization, AI Agents can help humans find mistakes faster, dig deeper into the causes, and come up with smarter solutions. It turns a slow, experience-based detective game into a fast, data-driven superpower, allowing hospitals to keep patients safer with less stress on their staff.

Technical Summary

Technical Summary: Application of AI Agent Technology in Quality Defect Root Cause Analysis of the Central Sterile Supply Department

Problem Statement
The Central Sterile Supply Department (CSSD) faces significant challenges in managing quality inspection defects. Traditional root cause analysis (RCA) methods, such as fishbone diagrams and the 5-Why technique, rely heavily on subjective human experience. This dependence leads to several operational pain points: prolonged analysis cycles, high labor costs due to the necessity of involving senior staff, delayed risk warnings, and insufficient analysis depth caused by cognitive limitations. Furthermore, traditional approaches often struggle to achieve dynamic root cause inference and fail to fully leverage historical data for predictive risk management. While AI has been applied to infection control predictive modeling, its application in assisting CSSD personnel with dynamic, real-time root cause analysis remains in a preliminary exploratory stage.

Methodology
This study employed a retrospective case simulation comparative study to evaluate an AI agent-assisted approach against traditional manual analysis.

• Data Source: 30 representative quality defect cases were selected from a tertiary Grade A hospital's CSSD records (January 2023–December 2025). Cases covered four categories: substandard instrument cleaning, packaging defects, non-compliant sterilization parameters, and instrument functional damage.
• Experimental Design:
  • Experimental Group (AI Agent): Utilized a specialized AI agent built on the "Doubao" large language model platform. The agent architecture consisted of three layers:
    1. Perception and Input: Structured case data entry by a QC specialist.
    2. Analysis and Inference Engine: Integrated multidimensional factor correlation analysis (People, Machine, Material, Method, Environment, Management), Knowledge Graphs (vectorized industry standards, SOPs, and equipment principles), Retrieval Augmented Generation (RAG) for retrieving historical cases and standards, and Chain-of-Thought (CoT) reasoning for logical deduction.
    3. Output and Validation: Generated structured reports including probable root causes, logical chains, and improvement suggestions.
  • Control Group (Traditional): Composed of experienced CSSD personnel (5+ years) using standard tools (fishbone diagrams, 5-Why).
• Evaluation Benchmark: A high-level expert panel (≥5 members) provided blind "Gold Standard" root cause lists and improvement plans.
• Metrics: Performance was measured across four dimensions:
  1. Quality: Root cause identification accuracy, analysis depth (5-level scale), comprehensiveness (dimension coverage), and logical chain integrity.
  2. Efficiency: Time per case, root causes identified per unit time, and time to first core root cause.
  3. Practicality: Feasibility and innovation of proposed measures.
  4. Process/User Experience: Process standardization, cognitive load (NASA-TLX), learning curve (novice performance), and user satisfaction.
• Statistical Analysis: Data were analyzed using SPSS 25.0/SPSSAU. Parametric tests (t-tests) and non-parametric tests (Mann-Whitney U) were used based on data distribution, with significance set at P<0.05.

Key Results
The AI agent-assisted group demonstrated statistically significant superiority over the traditional group in most metrics:

• Analysis Quality: The AI group achieved higher root cause identification accuracy (85.6% vs. 72.3%, P<0.001), deeper analysis (4.4 vs. 3.6 points, P<0.001), and greater comprehensiveness (5.2 vs. 4.1 dimensions, P<0.001). Logical chain integrity was also higher (4.2 vs. 3.9, P=0.045).
• Efficiency: The AI group significantly reduced analysis time (18.5 vs. 35.2 minutes, P<0.001) and identified the first core root cause much faster (8.3 vs. 19.6 minutes, P<0.001).
• Practicality: While the feasibility of measures was comparable between groups (P>0.05), the AI group generated significantly more innovative measures (1.8 vs. 0.7 items/case, P<0.001) and more systematic solutions (4.1 vs. 3.5 points, P=0.001).
• User Experience: The AI group showed 100% process standardization compliance (vs. 83.3%), significantly lower cognitive load (42.5 vs. 68.2 on NASA-TLX), and higher novice-to-expert result similarity (78.4% vs. 62.1%).

Key Contributions

1. Development of a Specialized AI Agent: The study constructed a modular AI agent integrating RAG, Knowledge Graphs, and CoT reasoning specifically tailored for CSSD quality control, moving beyond simple predictive modeling to dynamic root cause inference.
2. Empirical Validation of Efficiency and Depth: The research provides quantitative evidence that AI agents can significantly accelerate defect analysis while simultaneously increasing the depth and systematic nature of the findings, addressing the "cognitive bottleneck" of human analysts.
3. Democratization of Expertise: The study demonstrates that AI acts as a "cognitive scaffold," enabling novice users to achieve analysis results with high similarity to expert panels, thereby reducing reliance on senior staff experience and shortening training cycles.
4. Cost-Benefit Analysis: The paper highlights that while initial tool investment is required, the long-term benefits include reduced training costs, faster analysis times, and the accumulation of reusable organizational knowledge assets, contrasting with the limited scalability of traditional experience-based methods.

Significance and Claims
The paper claims that integrating AI agent technology into CSSD quality management represents a shift from experience-driven to data-intelligence-driven quality control. The study asserts that this approach effectively reduces defect response times from hours to minutes and enhances the consistency and traceability of analytical results.

The authors position the AI not as a replacement for human judgment but as a collaborative tool that externalizes expert logic. They argue that the ideal future model is a "human-machine symbiotic intelligent ecosystem" where AI handles foundational analysis and pattern recognition, freeing human experts to focus on strategic innovation, ethical deliberation, and value calibration. The study concludes that this technology has high application value for standardizing RCA processes, addressing workforce skill gaps, and enabling proactive, data-driven decision-making in medical quality control.