Negation's not solved: generalizability versus optimizability in clinical natural language processing

Abstract

A review of published work in clinical natural language processing (NLP) may suggest that the negation detection task has been "solved." This work proposes that an optimizable solution does not equal a generalizable solution. We introduce a new machine learning-based Polarity Module for detecting negation in clinical text, and extensively compare its performance across domains. Using four manually annotated corpora of clinical text, we show that negation detection performance suffers when there is no in-domain development (for manual methods) or training data (for machine learning-based methods). Various factors (e.g., annotation guidelines, named entity characteristics, the amount of data, and lexical and syntactic context) play a role in making generalizability difficult, but none completely explains the phenomenon. Furthermore, generalizability remains challenging because it is unclear whether to use a single source for accurate data, combine all sources into a single model, or apply domain adaptation methods. The most reliable means to improve negation detection is to manually annotate in-domain training data (or, perhaps, manually modify rules); this is a strategy for optimizing performance, rather than generalizing it. These results suggest a direction for future work in domain-adaptive and task-adaptive methods for clinical NLP.

Figure 1. The cTAKES Pipeline.

The SHARPn Polarity Module is an Attribute Discovery algorithm. Training and evaluations use gold standard NEs (skip NER).

Figure 2. Significance bands of model performance for each test corpus.

These are labeled with successive letters from right to left in Table 4.

Figure 3. Learning curve for i2b2 training data on various corpora.

For each proportion of the i2b2 corpus (x axis), the reported F-score (y axis) is an average of 5 randomly sampled runs.

Figure 4. The effect of named entity length (in number of words) on performance for each of 6 training configurations.

SHARP, MiPACQ, and i2b2 test sets are used for evaluation.

Figure 5. The effect of named entity semantic group on the F-score of 6 models.

SHARP, MiPACQ, and i2b2 test sets are used for evaluation.