Neural Network Medical Coding Using Hierarchical Semantic Weights

The patent replaces traditional mechanical string matching and rule-based techniques with a neural network system that uses learned semantic representations. The neural network processes adverse event citations by transforming them into vector embeddings and comparing them with coded event embeddings, enabling the system to capture semantic meaning rather than relying on exact string matches or predefined rules.

The patent changes the parameter space from discrete string matching to continuous vector space embeddings. By representing both adverse event citations and coded events as vectors in a high-dimensional space, the system can measure semantic similarity through distance metrics, allowing for more nuanced and accurate matching that captures subtle semantic differences between medical terms.

The patent performs preliminary action by pre-computing embeddings for all coded events in the medical dictionary and storing them in a reference database. When a new adverse event citation needs to be coded, the system only needs to compute the embedding for that single citation and compare it against the pre-computed embeddings, rather than traversing and evaluating all possible code nodes in the medical dictionary hierarchy.

The patent introduces vector embeddings as an intermediary representation layer between the raw adverse event citation text and the discrete medical codes. This embedding layer acts as a mediator that transforms the coding problem from a discrete search through all possible codes into a continuous similarity search, enabling efficient retrieval of the most appropriate codes without exhaustive traversal.

The patent implements self-service by training the neural network to automatically learn the complex mapping between adverse event citations and medical codes without requiring human experts to perform each coding task manually. The system uses training data annotated by experts to learn the coding patterns, then autonomously performs coding on new data, maintaining high accuracy while achieving scalable automation.

The patent incorporates feedback mechanisms where the neural network is trained on labeled data from expert-coded examples, using the expert annotations as ground truth to guide learning. The system continuously improves its coding accuracy by learning from these feedback signals, capturing the expertise of human coders in a reusable computational model that can process unlimited numbers of citations.