School of Medicine, University of Virginia, Charlottesville, Virginia.
Department of Public Health, University of Virginia, Charlottesville, Virginia.
Traffic Inj Prev. 2024;25(sup1):S25-S32. doi: 10.1080/15389588.2024.2356663. Epub 2024 Nov 1.
The injury severity classification based on the Abbreviated Injury Scale (AIS) provides information that allows for standardized comparisons for injury research. However, the majority of injury data is captured using the International Classification of Diseases (ICD), which lacks injury severity information. It has been shown that the encoder-decoder-based neural machine translation (NMT) model is more accurate than other methods for determining injury severity from ICD codes. The objectives of this project were to determine if feed-forward neural networks (FFNN) perform as well as NMT and to determine if direct estimation of injury severity is more accurate than using AIS codes as an intermediary (indirect method).
Patient data from the National Trauma Data Bank were used to develop and test the four models (NMT/Indirect, NMT/Direct, FFNN/Indirect, FFNN/Direct). There were 2,031,793 cases from 2017-2018 used to train and 1,091,792 cases from 2019 were used for testing. The primary outcome of interest was the percent of cases with the correct binary classification of Injury Severity Score (ISS) ≥16, using ISS values recorded in NTDB for benchmarking. The secondary outcome was the percent of predicted ISS exactly matching the recorded ISS.
The results show that indirect estimation through first converting to AIS using an NMT was the most accurate in predicting ISS ≥ 16 (94.0%), followed by direct estimation with FFNN (93.4%), direct estimation with NMT (93.1%), and then indirect estimation with FFNN (93.1%), with statistically significant differences in pairwise comparison. The rankings were the same when evaluating models based on exactly matches of ISS. Training times were similar for all models (range 11-14 h), but testing was much faster for FFNN models (GPU: 1-2 min) compared to the NMT models (GPU: 69-82 min).
The most accurate method for obtaining injury severity from ICD was NMT using AIS codes as an intermediary (indirect method), although all methods performed well. The indirect NMT model was the most resource intensive in terms of processing time. The optimal approach for researchers will be based on their needs and the computing resources available.
基于简明损伤定级（AIS）的损伤严重度分类提供了可用于标准化比较的信息，从而有助于损伤研究。然而，大多数损伤数据是使用国际疾病分类（ICD）捕获的，而 ICD 缺乏损伤严重度信息。已经表明，基于编码器-解码器的神经机器翻译（NMT）模型比其他方法更准确，可以根据 ICD 代码确定损伤严重度。本项目的目的是确定前馈神经网络（FFNN）是否与 NMT 一样有效，以及直接估计损伤严重度是否比使用 AIS 代码作为中介（间接方法）更准确。
使用国家创伤数据库（NTDB）中的患者数据来开发和测试四个模型（NMT/间接、NMT/直接、FFNN/间接、FFNN/直接）。使用 2017-2018 年的 2031793 例数据进行训练，使用 2019 年的 1091792 例数据进行测试。主要研究结果是使用 NTDB 记录的损伤严重度评分（ISS）值作为基准，正确分类 ISS≥16 的病例百分比。次要结果是预测的 ISS 与记录的 ISS 完全匹配的百分比。
结果表明，通过首先使用 NMT 将其转换为 AIS 进行间接估计是预测 ISS≥16 最准确的方法（94.0%），其次是使用 FFNN 进行直接估计（93.4%）、使用 NMT 进行直接估计（93.1%），然后是使用 FFNN 进行间接估计（93.1%），两两比较差异具有统计学意义。根据 ISS 完全匹配评估模型时，排名相同。所有模型的训练时间相似（范围为 11-14 小时），但 FFNN 模型的测试速度要快得多（GPU：1-2 分钟），而 NMT 模型的测试速度要慢得多（GPU：69-82 分钟）。
从 ICD 获得损伤严重度最准确的方法是使用 AIS 代码作为中介的 NMT（间接方法），尽管所有方法都表现良好。间接 NMT 模型在处理时间方面是最耗费资源的。研究人员的最佳方法将基于他们的需求和可用的计算资源。
