|Home | About | Journals | Submit | Contact Us | Français|
To the Editor:
Greenwald et al1 propose using free text in patient records to estimate hospital readmission risk. They use expert knowledge to identify 35 groups of phrases indicative of 30-day rehospitalization, and use 16 of these in logistic regression. We believe the use of natural language processing (NLP) for predicting rehospitalization is an interesting approach, and provide suggestions to improve the model.
The proposed terms are all n-grams (n≤4) and therefore a subset of simpler bag-of-words,2 which can be extracted with lighter expert workload. Grouping terms to create variables can be done automatically using topic modeling.3 Taking context into account and normalizing abbreviations and word variants, as discussed by the authors, can be done using common-off-the-shelf software such as cTAKES.4 Graph modeling is another document representation for classification that has been shown to have good interpretability by experts.5
The distortion of the coefficients in table 3 and the modest improvements over the baseline suggest that the variables may share the same information. The Pearson correlation coefficients of all variables would help determine whether this is the case.
Another concern is that the proposed method is only compared with a baseline of prior hospitalizations. To evaluate the added value of the proposed variables, a stronger baseline could use all available structured data in the patient records that have been shown to have predictive value, that is, age, sex, comorbidity index.6 This also contributes to measuring the true effect of the proposed variables when adjusting for potential confounders.
The study of rehospitalization risks presents an excellent opportunity to assess the contribution of NLP to predicting important clinical outcomes. With this letter we want to encourage a more thorough evaluation of NLP methods toward this goal.
On behalf of the Methods in Research on Research (MiRoR) project.
Supported by the European Union’s Horizon 2020 research and innovation programme under the Marie Sklodowska-Curie grant agreement no. 676207.
The authors declare no conflict of interest.