Dataset Information

Predicting Postoperative Mortality With Deep Neural Networks and Natural Language Processing: Model Development and Validation.

ABSTRACT:

Background

Machine learning (ML) achieves better predictions of postoperative mortality than previous prediction tools. Free-text descriptions of the preoperative diagnosis and the planned procedure are available preoperatively. Because reading these descriptions helps anesthesiologists evaluate the risk of the surgery, we hypothesized that deep learning (DL) models with unstructured text could improve postoperative mortality prediction. However, it is challenging to extract meaningful concept embeddings from this unstructured clinical text.

Objective

This study aims to develop a fusion DL model containing structured and unstructured features to predict the in-hospital 30-day postoperative mortality before surgery. ML models for predicting postoperative mortality using preoperative data with or without free clinical text were assessed.

Methods

We retrospectively collected preoperative anesthesia assessments, surgical information, and discharge summaries of patients undergoing general and neuraxial anesthesia from electronic health records (EHRs) from 2016 to 2020. We first compared the deep neural network (DNN) with other models using the same input features to demonstrate effectiveness. Then, we combined the DNN model with bidirectional encoder representations from transformers (BERT) to extract information from clinical texts. The effects of adding text information on the model performance were compared using the area under the receiver operating characteristic curve (AUROC) and the area under the precision-recall curve (AUPRC). Statistical significance was evaluated using P<.05.

Results

The final cohort contained 121,313 patients who underwent surgeries. A total of 1562 (1.29%) patients died within 30 days of surgery. Our BERT-DNN model achieved the highest AUROC (0.964, 95% CI 0.961-0.967) and AUPRC (0.336, 95% CI 0.276-0.402). The AUROC of the BERT-DNN was significantly higher compared to logistic regression (AUROC=0.952, 95% CI 0.949-0.955) and the American Society of Anesthesiologist Physical Status (ASAPS AUROC=0.892, 95% CI 0.887-0.896) but not significantly higher compared to the DNN (AUROC=0.959, 95% CI 0.956-0.962) and the random forest (AUROC=0.961, 95% CI 0.958-0.964). The AUPRC of the BERT-DNN was significantly higher compared to the DNN (AUPRC=0.319, 95% CI 0.260-0.384), the random forest (AUPRC=0.296, 95% CI 0.239-0.360), logistic regression (AUPRC=0.276, 95% CI 0.220-0.339), and the ASAPS (AUPRC=0.149, 95% CI 0.107-0.203).

Conclusions

Our BERT-DNN model has an AUPRC significantly higher compared to previously proposed models using no text and an AUROC significantly higher compared to logistic regression and the ASAPS. This technique helps identify patients with higher risk from the surgical description text in EHRs.

SUBMITTER: Chen PF

PROVIDER: S-EPMC9131148 | biostudies-literature |

REPOSITORIES: biostudies-literature

ACCESS DATA

Similar Datasets

Project description:PurposeAdvances in artificial intelligence have produced a few predictive models in glaucoma, including a logistic regression model predicting glaucoma progression to surgery. However, uncertainty exists regarding how to integrate the wealth of information in free-text clinical notes. The purpose of this study was to predict glaucoma progression requiring surgery using deep learning (DL) approaches on data from electronic health records (EHRs), including features from structured clinical data and from natural language processing of clinical free-text notes.DesignDevelopment of DL predictive model in an observational cohort.ParticipantsAdult patients with glaucoma at a single center treated from 2008 through 2020.MethodsOphthalmology clinical notes of patients with glaucoma were identified from EHRs. Available structured data included patient demographic information, diagnosis codes, prior surgeries, and clinical information including intraocular pressure, visual acuity, and central corneal thickness. In addition, words from patients' first 120 days of notes were mapped to ophthalmology domain-specific neural word embeddings trained on PubMed ophthalmology abstracts. Word embeddings and structured clinical data were used as inputs to DL models to predict subsequent glaucoma surgery.Main outcome measuresEvaluation metrics included area under the receiver operating characteristic curve (AUC) and F1 score, the harmonic mean of positive predictive value, and sensitivity on a held-out test set.ResultsSeven hundred forty-eight of 4512 patients with glaucoma underwent surgery. The model that incorporated both structured clinical features as well as input features from clinical notes achieved an AUC of 73% and F1 of 40%, compared with only structured clinical features, (AUC, 66%; F1, 34%) and only clinical free-text features (AUC, 70%; F1, 42%). All models outperformed predictions from a glaucoma specialist's review of clinical notes (F1, 29.5%).ConclusionsWe can successfully predict which patients with glaucoma will need surgery using DL models on EHRs unstructured text. Models incorporating free-text data outperformed those using only structured inputs. Future predictive models using EHRs should make use of information from within clinical free-text notes to improve predictive performance. Additional research is needed to investigate optimal methods of incorporating imaging data into future predictive models as well.

Project description:Early identification of individuals at risk of developing Alzheimer's disease (AD) dementia is important for developing disease-modifying therapies. In this study, given multimodal AD markers and clinical diagnosis of an individual from one or more timepoints, we seek to predict the clinical diagnosis, cognition and ventricular volume of the individual for every month (indefinitely) into the future. We proposed and applied a minimal recurrent neural network (minimalRNN) model to data from The Alzheimer's Disease Prediction Of Longitudinal Evolution (TADPOLE) challenge, comprising longitudinal data of 1677 participants (Marinescu et al., 2018) from the Alzheimer's Disease Neuroimaging Initiative (ADNI). We compared the performance of the minimalRNN model and four baseline algorithms up to 6 years into the future. Most previous work on predicting AD progression ignore the issue of missing data, which is a prevalent issue in longitudinal data. Here, we explored three different strategies to handle missing data. Two of the strategies treated the missing data as a "preprocessing" issue, by imputing the missing data using the previous timepoint ("forward filling") or linear interpolation ("linear filling). The third strategy utilized the minimalRNN model itself to fill in the missing data both during training and testing ("model filling"). Our analyses suggest that the minimalRNN with "model filling" compared favorably with baseline algorithms, including support vector machine/regression, linear state space (LSS) model, and long short-term memory (LSTM) model. Importantly, although the training procedure utilized longitudinal data, we found that the trained minimalRNN model exhibited similar performance, when using only 1 input timepoint or 4 input timepoints, suggesting that our approach might work well with just cross-sectional data. An earlier version of our approach was ranked 5th (out of 53 entries) in the TADPOLE challenge in 2019. The current approach is ranked 2nd out of 63 entries as of June 3rd, 2020.

Project description:The ability to predict psychiatric readmission would facilitate the development of interventions to reduce this risk, a major driver of psychiatric health-care costs. The symptoms or characteristics of illness course necessary to develop reliable predictors are not available in coded billing data, but may be present in narrative electronic health record (EHR) discharge summaries. We identified a cohort of individuals admitted to a psychiatric inpatient unit between 1994 and 2012 with a principal diagnosis of major depressive disorder, and extracted inpatient psychiatric discharge narrative notes. Using these data, we trained a 75-topic Latent Dirichlet Allocation (LDA) model, a form of natural language processing, which identifies groups of words associated with topics discussed in a document collection. The cohort was randomly split to derive a training (70%) and testing (30%) data set, and we trained separate support vector machine models for baseline clinical features alone, baseline features plus common individual words and the above plus topics identified from the 75-topic LDA model. Of 4687 patients with inpatient discharge summaries, 470 were readmitted within 30 days. The 75-topic LDA model included topics linked to psychiatric symptoms (suicide, severe depression, anxiety, trauma, eating/weight and panic) and major depressive disorder comorbidities (infection, postpartum, brain tumor, diarrhea and pulmonary disease). By including LDA topics, prediction of readmission, as measured by area under receiver-operating characteristic curves in the testing data set, was improved from baseline (area under the curve 0.618) to baseline+1000 words (0.682) to baseline+75 topics (0.784). Inclusion of topics derived from narrative notes allows more accurate discrimination of individuals at high risk for psychiatric readmission in this cohort. Topic modeling and related approaches offer the potential to improve prediction using EHRs, if generalizability can be established in other clinical cohorts.

Project description:Many recent advances in artificial intelligence (AI) are rooted in visual neuroscience. However, ideas from more complicated paradigms like decision-making are less used. Although automated decision-making systems are ubiquitous (driverless cars, pilot support systems, medical diagnosis algorithms etc.), achieving human-level performance in decision making tasks is still a challenge. At the same time, these tasks that are hard for AI are easy for humans. Thus, understanding human brain dynamics during these decision-making tasks and modeling them using deep neural networks could improve AI performance. Here we modelled some of the complex neural interactions during a sensorimotor decision making task. We investigated how brain dynamics flexibly represented and distinguished between sensory processing and categorization in two sensory domains: motion direction and color. We used two different approaches for understanding neural representations. We compared brain responses to 1) the geometry of a sensory or category domain (domain selectivity) and 2) predictions from deep neural networks (computation selectivity). Both approaches gave us similar results. This confirmed the validity of our analyses. Using the first approach, we found that neural representations changed depending on context. We then trained deep recurrent neural networks to perform the same tasks as the animals. Using the second approach, we found that computations in different brain areas also changed flexibly depending on context. Color computations appeared to rely more on sensory processing, while motion computations more on abstract categories. Overall, our results shed light to the biological basis of categorization and differences in selectivity and computations in different brain areas. They also suggest a way for studying sensory and categorical representations in the brain: compare brain responses to both a behavioral model and a deep neural network and test if they give similar results.