Concept Extraction Exercise using NotebookLM (Google)

Forty of this study's 100 sample malpractice claims generated by Microsoft Copilot were uploaded into NotebookLM for analysis. The first task asked the AI to extract up to three misadventures per case without using a predefined schema.

A second phase introduced a structured list of misadventure codes, prompting the AI to align its outputs with this taxonomy. In addition, NotebookLM was asked to apply standardised clinical coding, assigning ICD‑10 disease codes and OPCS‑4 treatment codes at the three‑character level. This allowed comparison between free‑form concept extraction, taxonomy‑based classification, and recognised clinical coding systems.

NotebookLM generally captured the key misadventures and even highlighted recurring systemic errors across cases. While occasional inconsistencies appeared (e.g., use of four‑digit ICD‑10 codes instead of three), overall results were judged accurate enough to demonstrate practical value.

For organisations, this technology may serve as a validation tool for existing coding teams or as a starting point for structured data capture. As always, sensitive information should only be submitted with appropriate safeguards, including de‑identification and compliance with data protection standards.

The table below consolidates the outputs from the two NotebookLM exercises—concept extraction and structured clinical coding—into one view. Each row shows the AI’s summary of misadventures alongside the assigned misadventure, disease, and treatment codes. These results are presented here not as part of the site’s own analytical method, but to illustrate that large‑scale, readily available LLM tools can generate alternative tabular outputs. By placing them together, readers can see how such systems operate in practice, while recognising that the site’s preferred approach remains the smaller‑scale, self‑contained analytical method described elsewhere.