A Linear Least Squares Fit Mapping Method For Information Retrieval From Natural Language Texts

Yiming Yang, Christopher G. ChuteCOLING 1992 · OpenReview

Abstract

In medical informatics, a significant gap exists between the natural language used by physicians in clinical records and the formal canonical terms required for database organization and retrieval. This paper presents a Linear Least Squares Fit (LLSF) method to automatically map free-text diagnoses to formal codes (such as ICD-9-CM). By treating the mapping as a translation problem between two "languages" and solving it through linear algebra, the LLSF approach achieves high accuracy, significantly outperforming traditional string-matching and statistical weighting techniques.

Introduction

Medical databases often rely on standardized coding systems like ICD-9-CM to organize patient information. However, the data entered by physicians is typically "free-text"—messy, natural language that varies significantly from formal terminology. Traditional information retrieval methods struggle with this mapping for two primary reasons:

  1. Synonymy (The "Too Little" Problem): String matching cannot identify that "high blood pressure" and "hypertension" are equivalent because they share no common characters or words.
  2. Polysemy and Noise (The "Too Many" Problem): Simple keyword searches for "high blood pressure" may incorrectly match unrelated terms like "high head at term" or "fetal blood loss" due to shared common words.

While manual thesauruses can bridge this gap, they are expensive to build and difficult to maintain across diverse medical domains. The LLSF method offers a numerical, algorithmic solution that learns these associations directly from existing expert-coded data.

The LLSF Methodology

The authors propose treating the mapping as a linear transformation. The problem is modeled as a translation between a source language (physician notes) and a target language (canonical terms).

Matrix Representation and Optimization

The system represents a collection of known matches as two matrices. Matrix A contains the source texts, where rows represent words and columns represent individual diagnoses. Matrix B contains the corresponding canonical terms. The goal is to find a mapping matrix W such that:

$$WA \approx B$$

This matrix W represents the weighted connections between every word in the source vocabulary and every word in the target vocabulary. To find the optimal W even in the presence of inconsistent or massive data, the researchers utilize Singular Value Decomposition (SVD), a robust linear algebra technique that identifies the "best fit" solution.

Figure 1: Matrix representation of source and target terms and the resulting LLSF solution matrix W.

Retrieval Process

When a new query is received, it is converted into a vector and multiplied by the pre-computed matrix W. This projects the query into the "canonical space." The system then compares this projected vector against the list of official terms using cosine similarity to find the most relevant matches.

Experimental Setup and Data

The study utilized a real-world dataset from the Mayo Clinic patient record archive, specifically focusing on cardiovascular diseases. The dataset included 6,913 pairs of diagnoses and their expert-assigned canonical terms from the ICD-9-CM classification system. To ensure the model's generalizability, the data was split into an "odd-half" for training and an "even-half" for testing.

Results and Performance

The LLSF method demonstrated a dramatic improvement over traditional benchmarks. While traditional string matching and statistical weighting (SMART) hovered around 15-21% recall for the top choice, LLSF achieved an 84% recall rate.

Figure 3: Performance comparison of LLSF against traditional string matching and statistical weighting methods.

The authors noted that only 20% of the words in physician queries actually appeared in the official ICD-9 terms. Traditional methods, which rely on surface-level word overlaps, effectively ignored 80% of the available information—information that LLSF successfully captured through its learned mapping.

Key Insights and Discussion

  • Context Sensitivity: Unlike simple word-weighting schemes, LLSF handles context through negative weights. For instance, it can learn to cancel out ambiguous matches for "high grade" when it appears alongside "carotid ulceration" rather than cancer-related terms.
  • Implicit Knowledge Capture: The system automatically "discovers" synonyms. If "glioma" is consistently mapped to "malignant neoplasm" in the training data, the matrix W develops strong positive connections between those terms without manual rule definition.
  • Limitations: While powerful, the model assumes a linear relationship between words. More complex, non-linear relationships might eventually require neural network models. Additionally, while retrieval is fast (0.45 seconds), the initial computation of the SVD for very large datasets can be computationally intensive.

In conclusion, the LLSF method provides a robust, mathematically clear framework for bridging the gap between natural language and formal medical coding, offering a scalable alternative to manual thesaurus construction.

Summary updated

BibTeX
@inproceedings{yang1992linear,
  author = {Yiming Yang and Christopher G. Chute},
  title = {{A Linear Least Squares Fit Mapping Method For Information Retrieval From Natural Language Texts}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1992},
  url = {https://openreview.net/forum?id=4CmCxcrLzzQ},
}
Figure