DescriptionAI-based recommendation systems are widely utilized in different fields including movies, music, news, social tags and products in general. Such systems may help reduce medical team errors and improve patient outcomes in treatment processes (e.g., trauma resuscitation, surgical processes) by extracting knowledge from historic data and providing online recommendations. We developed data-driven process recommender systems for trauma resuscitations process based on different models. This thesis includes three main topics: (1) process data augmentation algorithms; (2) two intention mining models; and (3) two process recommender systems. Topic (1) and (2) were developed for improving the performance of recommender systems (Topic (3)).
Our process data was collected manually by medical experts reviewing the recorded videos. The data collection was labor intensive and we coded 123 trauma patient records in the past four years. Because of the small size of our dataset, we attempted to augment it by generating synthetic data. We developed two synthetic data generators to augment our dataset: (1) alignment-based process data generator and (2) sequential generative adversarial network. Both of them can generate large amounts of semi-synthetic process data that has similar characteristics with those of real-world process data.
We used intention mining models to discover the relationship between observed treatment activities and medical team’s underlying intentions. By identifying medical team’s intentions, we are able to generate accurate recommendations. We developed
two different intention mining algorithms, one based on Hidden Markov Models and the other based on Seq2seq models.
Last, we designed the process recommendation systems using two different models, (1) Hierarchical Hidden Markov Model (HHMM) and (2) Long Short-Term Memory (LSTM). The HHMM-based recommender system utilizes the intention mining algorithm to estimate the medical team’s current intention, and then provides the process recommendation identified in that intention category. On the other hand, the LSTM-based recommender system learns the relationships from different processes. And also, the LSTM model was modified to deal with both environmental (i.e., patient demographics) and behavioral (i.e., preceding treatment activities) contextual information. To provide the process recommendation, the LSTM is iterated over the previous process trace, and uses the most likely activity as the next-step recommending process. For HHMM-based recommender system, we achieved top-1 accuracy at 34.4% and top-5 accuracy 56.9% over 102 kinds of activities. The LSTM-based recommender system showed a higher top-1 accuracy at 39.9% and top-5 accuracy 65.5%. The experimental results indicated both of out recommender systems (HHMM & LSTM) outperforms baseline models in recommendation accuracy, demonstrating the feasibility of our context-aware process recommendation systems for complex real-world medical processes.