Background: Adverse drug reactions (ADRs) are an important concern in the medication process and can pose a substantial economic burden for patients and hospitals. Because of the limitations of clinical trials, it is difficult to identify all possible ADRs of a drug before it is marketed. We developed a new model based on data mining technology to predict potential ADRs based on available drug data. Method: Based on the Word2Vec model in Nature Language Processing, we propose a new knowledge graph embedding method that embeds drugs and ADRs into their respective vectors and builds a logistic regression classification model to predict whether a given drug will have ADRs. Result: First, a new knowledge graph embedding method was proposed, and comparison with similar studies showed that our model not only had high prediction accuracy but also was simpler in model structure. In our experiments, the AUC of the classification model reached a maximum of 0.87, and the mean AUC was 0.863. Conclusion: In this paper, we introduce a new method to embed knowledge graph to vectorize drugs and ADRs, then use a logistic regression classification model to predict whether there is a causal relationship between them. The experiment showed that the use of knowledge graph embedding can effectively encode drugs and ADRs. And the proposed ADRs prediction system is also very effective.
Computer-assisted clinical coding (CAC) based on automated coding algorithms has been expected to improve the International Classification of Disease, tenth version (ICD-10) coding quality and productivity, whereas studies oriented to primary diagnosis auto-coding are limited in the Chinese context. This study aims at developing a machine learning (ML) model for automated primary diagnosis ICD-10 coding. A total of 71,709 admissions in Fuwai hospital were included to carry out this study, corresponding to 168 primary diagnosis ICD-10 codes. Based on clinical implications, two feature engineering methods were used to process discharge diagnosis and procedure texts into sequential features and sequential grouping features respectively by which two kinds of models were built and compared. One baseline model using one-hot encoding features was considered. Light Gradient Boosting Machine (LightGBM) was adopted as the classifier, and grid search and cross-validation were used to select the optimal hyperparameters. SHapley Additive exPlanations (SHAP) values were applied to give the interpretability of models. Our best prediction model was developed based on sequential grouping features. It showed good performance in the test phase with accuracy and macro-averaged F1 (Macro-F1) of 95.2% and 88.3% respectively. The comparison of the models demonstrated the effectiveness of the sequential information and the grouping strategy in boosting model performance (P-value < 0.01). Subgroup analysis of the best model on each individual code manifested that 91.1% of the codes achieved the F1 over 70.0%. Our model has been demonstrated its effectiveness for automated primary diagnosis coding in the Chinese context and its results are interpretable. Hence, it has the potential to assist clinical coders to improve coding efficiency and quality in Chinese inpatient settings.
Background: Secondary hypertension is a kind of hypertension with a definite etiology and may be cured. Patients with suspected secondary hypertension can benefit from timely detection and treatment and, conversely, will have a higher risk of morbidity and mortality than those with primary hypertension. Objective: The aim of this study was to develop and validate machine learning (ML) prediction models of common etiologies in patients with suspected secondary hypertension. Methods: The analyzed data set was retrospectively extracted from electronic medical records of patients discharged from Fuwai Hospital between January 1, 2016, and June 30, 2019. A total of 7532 unique patients were included and divided into 2 data sets by time: 6302 patients in 2016-2018 as the training data set for model building and 1230 patients in 2019 as the validation data set for further evaluation. Extreme Gradient Boosting (XGBoost) was adopted to develop 5 models to predict 4 etiologies of secondary hypertension and occurrence of any of them (named as composite outcome), including renovascular hypertension (RVH), primary aldosteronism (PA), thyroid dysfunction, and aortic stenosis. Both univariate logistic analysis and Gini Impurity were used for feature selection. Grid search and 10-fold cross-validation were used to select the optimal hyperparameters for each model. Results: Validation of the composite outcome prediction model showed good performance with an area under the receiver-operating characteristic curve (AUC) of 0.924 in the validation data set, while the 4 prediction models of RVH, PA, thyroid dysfunction, and aortic stenosis achieved AUC of 0.938, 0.965, 0.959, and 0.946, respectively, in the validation data set. A total of 79 clinical indicators were identified in all and finally used in our prediction models. The result of subgroup analysis on the composite outcome prediction model demonstrated high discrimination with AUCs all higher than 0.890 among all age groups of adults. Conclusions: The ML prediction models in this study showed good performance in detecting 4 etiologies of patients with suspected secondary hypertension; thus, they may potentially facilitate clinical diagnosis decision making of secondary hypertension in an intelligent way.
12 January 2022 in BMC Medical Informatics and Decision Making
Background: Automated ICD coding on medical texts via machine learning has been a hot topic. Related studies from medical field heavily relies on conventional bag-of-words (BoW) as the feature extraction method, and do not commonly use more complicated methods, such as word2vec (W2V) and large pretrained models likeBERT. This study aimed at uncovering the most effective feature extraction methods for coding models by comparingBoW,W2VandBERTvariants.Methods: We experimented with a Chinese dataset from Fuwai Hospital, which contains 6947 records and 1532 unique ICD codes, and a public Spanish dataset, which contains 1000 records and 2557 unique ICD codes. We designed coding tasks with different code frequency thresholds (denoted as $$f_s$$ ), with a lower threshold indicating a more complex task. Using traditional classifiers, we comparedBoW,W2VandBERTvariants on accomplishing these coding tasks.Results: When $$f_s$$ was equal to or greater than 140 for Fuwai dataset, and 60 for the Spanish dataset, theBERTvariants with the whole network fine-tuned was the best method, leading to aMicro-F1 of 93.9% for Fuwai data when $$f_s=200$$ , and aMicro-F1 of 85.41% for the Spanish dataset when $$f_s=180$$ . When $$f_s$$ fell below 140 for Fuwai dataset, and 60 for the Spanish dataset,BoWturned out to be the best, leading to aMicro-F1 of 83% for Fuwai dataset when $$f_s=20$$ , and aMicro-F1 of 39.1% for the Spanish dataset when $$f_s=20$$ . Our experiments also showed that both theBERTvariants andBoWpossessed good interpretability, which is important for medical applications of coding models.Conclusions: This study shed light on building promising machine learning models for automated ICD coding by revealing the most effective feature extraction methods. Concretely, our results indicated that fine-tuning the whole network of theBERTvariants was the optimal method for tasks covering only frequent codes, especially codes that represented unspecified diseases, whileBoWwas the best for tasks involving both frequent and infrequent codes. The frequency threshold where the best-performing method varied differed between different datasets due to factors like language and codeset.