Natural language processing for prediction of readmission in posterior lumbar fusion patients: which free-text notes have the most utility?

Natural language processing in low back pain and spine diseases: A systematic review

“…Prediction tasks focus on predicting post-operative outcomes. In their first paper, Karhade et al [ 37 ], they attempted to identify required re-operations due to wound infections arising after lumbar discectomy, while in a subsequent study [ 38 ] they identified unplanned re-admissions of patients who underwent posterior lumbar fusion. Both the tasks were intended to refer to a period within 90 days.…”

“…Other examples include operative notes, obviously used for the intra-operative tasks [ 31 – 34 ], and post-operative ones too [ 37 , 38 ]. Furthermore, the article from Karhade et al [ 38 ], compared different kinds of clinical notes, including discharge summaries [ 22 ], and physicians and nursing notes.…”

“…Other examples include operative notes, obviously used for the intra-operative tasks [ 31 – 34 ], and post-operative ones too [ 37 , 38 ]. Furthermore, the article from Karhade et al [ 38 ], compared different kinds of clinical notes, including discharge summaries [ 22 ], and physicians and nursing notes. With the exception of [ 36 ], in which Galbusera et al exploited notes in Italian, all other studies referred to notes written in English language.…”

“…In both cases, the model was implemented together with a Least Absolute Shrinkage and Selection Operator (LASSO) regularization. The latter was particularly employed by Karhade et al in several tasks [ 31 – 33 , 37 , 38 ]. Another used algorithm was the Support Vector Machine (SVM), employed in [ 23 ] to identify clues of axSpA and in [ 24 ] both to directly identify axSpA and to extract a feature for a multimodal random forest.…”

“…While discrimination measures the predictor’s ability to separate patients with different responses, calibration captures the degree to which its numerical predictions match the outcomes [ 47 , 48 ]. In particular, some of the analyzed works reported intercept and slope measures [ 31 – 33 , 37 , 38 ] to assess the miscalibration of the system. Specifically, a positive/negative calibration intercept assesses the over-/under-estimation of the predictions, while a calibration slope evaluates the spread of the predictions; a slope greater/lower than 1 would indicate that the predictions are too moderate/extreme.…”