RESUMO

BACKGROUND: In AFP-negative hepatocellular carcinoma patients, markers for predicting tumor progression or prognosis are limited. Therefore, our objective is to establish an optimal predicet model for this subset of patients, utilizing interpretable methods to enhance the accuracy of HCC prognosis prediction. METHODS: We recruited a total of 508 AFP-negative HCC patients in this study, modeling with randomly divided training set and validated with validation set. At the same time, 86 patients treated in different time periods were used as internal validation. After comparing the cox model with the random forest model based on Lasso regression, we have chosen the former to build our model. This model has been interpreted with SHAP values and validated using ROC, DCA. Additionally, we have reconfirmed the model's effectiveness by employing an internal validation set of independent periods. Subsequently, we have established a risk stratification system. RESULTS: The AUC values of the Lasso-Cox model at 1, 2, and 3 years were 0.807, 0.846, and 0.803, and the AUC values of the Lasso-RSF model at 1, 2, and 3 years were 0.783, 0.829, and 0.776. Lasso-Cox model was finally used to predict the prognosis of AFP-negative HCC patients in this study. And BCLC stage, gamma-glutamyl transferase (GGT), diameter of tumor, lung metastases (LM), albumin (ALB), alkaline phosphatase (ALP), and the number of tumors were included in the model. The validation set and the separate internal validation set both indicate that the model is stable and accurate. Using risk factors to establish risk stratification, we observed that the survival time of the low-risk group, the middle-risk group, and the high-risk group decreased gradually, with significant differences among the three groups. CONCLUSION: The Lasso-Cox model based on AFP-negative HCC showed good predictive performance for liver cancer. SHAP explained the model for further clinical application.

RESUMO

Early and accurate detection of side effects is critical for the clinical success of drugs under development. Here, we aim to predict unknown side effects for drugs with a small number of side effects identified in randomized controlled clinical trials. Our machine learning framework, the geometric self-expressive model (GSEM), learns globally optimal self-representations for drugs and side effects from pharmacological graph networks. We show the usefulness of the GSEM on 505 therapeutically diverse drugs and 904 side effects from multiple human physiological systems. Here, we also show a data integration strategy that could be adopted to improve the ability of side effect prediction models to identify unknown side effects that might only appear after the drug enters the market.

Assuntos

RESUMO

In the last five years, the inclusion of Deep Learning algorithms in prognostics and health management (PHM) has led to a performance increase in diagnostics, prognostics, and anomaly detection. However, the lack of interpretability of these models results in resistance towards their deployment. Deep Learning-based models fall within the accuracy/interpretability tradeoff, which means that their complexity leads to high performance levels but lacks interpretability. This work aims at addressing this tradeoff by proposing a technique for feature selection embedded in deep neural networks that uses a feature selection (FS) layer trained with the rest of the network to evaluate the input features' importance. The importance values are used to determine which will be considered for deployment of a PHM model. For comparison with other techniques, this paper introduces a new metric called ranking quality score (RQS), that measures how performance evolves while following the corresponding ranking. The proposed framework is exemplified with three case studies involving health state diagnostics and prognostics and remaining useful life prediction. Results show that the proposed technique achieves higher RQS than the compared techniques, while maintaining the same performance level when compared to the same model but without an FS layer.

Assuntos

RESUMO

BACKGROUND: Displeasure with the functionality of clinical decision support systems (CDSSs) is considered the primary challenge in CDSS development. A major difficulty in CDSS design is matching the functionality to the desired and actual clinical workflow. Computer-interpretable guidelines (CIGs) are used to formalize medical knowledge in clinical practice guidelines (CPGs) in a computable language. However, existing CIG frameworks require a specific interpreter for each CIG language, hindering the ease of implementation and interoperability. OBJECTIVE: This paper aims to describe a different approach to the representation of clinical knowledge and data. We intended to change the clinician's perception of a CDSS with sufficient expressivity of the representation while maintaining a small communication and software footprint for both a web application and a mobile app. This approach was originally intended to create a readable and minimal syntax for a web CDSS and future mobile app for antenatal care guidelines with improved human-computer interaction and enhanced usability by aligning the system behavior with clinical workflow. METHODS: We designed and implemented an architecture design for our CDSS, which uses the model-view-controller (MVC) architecture and a knowledge engine in the MVC architecture based on XML. The knowledge engine design also integrated the requirement of matching clinical care workflow that was desired in the CDSS. For this component of the design task, we used a work ontology analysis of the CPGs for antenatal care in our particular target clinical settings. RESULTS: In comparison to other common CIGs used for CDSSs, our XML approach can be used to take advantage of the flexible format of XML to facilitate the electronic sharing of structured data. More importantly, we can take advantage of its flexibility to standardize CIG structure design in a low-level specification language that is ubiquitous, universal, computationally efficient, integrable with web technologies, and human readable. CONCLUSIONS: Our knowledge representation framework incorporates fundamental elements of other CIGs used in CDSSs in medicine and proved adequate to encode a number of antenatal health care CPGs and their associated clinical workflows. The framework appears general enough to be used with other CPGs in medicine. XML proved to be a language expressive enough to describe planning problems in a computable form and restrictive and expressive enough to implement in a clinical system. It can also be effective for mobile apps, where intermittent communication requires a small footprint and an autonomous app. This approach can be used to incorporate overlapping capabilities of more specialized CIGs in medicine.

RESUMO

Although many authors have highlighted the importance of predicting people's health costs to improve healthcare budget management, most of them do not address the frequent need to know the reasons behind this prediction, i.e., knowing the factors that influence this prediction. This knowledge allows avoiding arbitrariness or people's discrimination. However, many times the black box methods (that is, those that do not allow this analysis, e.g., methods based on deep learning techniques) are more accurate than those that allow an interpretation of the results. For this reason, in this work, we intend to develop a method that can achieve similar returns as those obtained with black box methods for the problem of predicting health costs, but at the same time it allows the interpretation of the results. This interpretable regression method is based on the Dempster-Shafer theory using Evidential Regression (EVREG) and a discount function based on the contribution of each dimension. The method "learns" the optimal weights for each feature using a gradient descent technique. The method also uses the nearest k-neighbor algorithm to accelerate calculations. It is possible to select the most relevant features for predicting a patient's health care costs using this approach and the transparency of the Evidential Regression model. We can obtain a reason for a prediction with a k-NN approach. We used the Japanese health records at Tsuyama Chuo Hospital to test our method, which included medical examinations, test results, and billing information from 2013 to 2018. We compared our model to methods based on an Artificial Neural Network, Gradient Boosting, Regression Tree and Weighted k-Nearest Neighbors. Our results showed that our transparent model performed like the Artificial Neural Network and Gradient Boosting with an R2 of 0.44.

Assuntos

RESUMO

BACKGROUND: The distribution of printed materials is the most frequently used strategy to disseminate and implement clinical practice guidelines, although several studies have shown that the effectiveness of this approach is modest at best. Nevertheless, there is insufficient evidence to support the use of other strategies. Recent research has shown that the use of computerized decision support presents a promising approach to address some aspects of this problem. OBJECTIVE: The aim of this study is to provide qualitative evidence on the potential effect of mobile decision support systems to facilitate the implementation of evidence-based recommendations included in clinical practice guidelines. METHODS: We will conduct a qualitative study with two arms to compare the experience of primary care physicians while they try to implement an evidence-based recommendation in their clinical practice. In the first arm, we will provide participants with a printout of the guideline article containing the recommendation, while in the second arm, we will provide participants with a mobile app developed after formalizing the recommendation text into a clinical algorithm. Data will be collected using semistructured and open interviews to explore aspects of behavioral change and technology acceptance involved in the implementation process. The analysis will be comprised of two phases. During the first phase, we will conduct a template analysis to identify barriers and facilitators in each scenario. Then, during the second phase, we will contrast the findings from each arm to propose hypotheses about the potential impact of the system. RESULTS: We have formalized the narrative in the recommendation into a clinical algorithm and have developed a mobile app. Data collection is expected to occur during 2018, with the first phase of analysis running in parallel. The second phase is scheduled to conclude in July 2019. CONCLUSIONS: Our study will further the understanding of the role of mobile decision support systems in the implementation of clinical practice guidelines. Furthermore, we will provide qualitative evidence to aid decisions made by low- and middle-income countries' ministries of health about investments in these technologies.