Interpretation Algorithm Research Articles

Identifying biotic stress-associated molecular markers in wheat using differential gene expression and machine learning techniques

Wheat is an important crop for global food security and a key crop for many developing countries. Thanks to next-generation sequencing (NGS) technologies, researchers can analyze the transcriptome of wheat and reveal differentially expressed genes (DEGs) responsible for essential agronomic traits and biotic stress tolerance. In addition, machine learning (ML) methods have opened new avenues to detect patterns in expression data and make predictions or decisions based on these patterns. We used both techniques to identify potential molecular markers in wheat associated with biotic stress in six gene expression studies conducted to investigate powdery mildew, blast fungus, rust, fly larval infection, greenbug aphid, and Stagonospora nodorum infections. A total of 24,152 threshold genes were collected from different studies, with the highest threshold being 7580 genes and the lowest being 1073 genes. The study identified several genes that were differentially expressed in all comparisons and genes that were present in only one data set. The study also discussed the possible role of certain genes in plant resistance. The Ta-TLP, HBP-1, WRKY, PPO, and glucan endo-1,3-beta-glucosidase genes were selected by the interpretable model-agnostic explanation algorithm as the most important genes known to play a significant role in resistance to biotic stress. Our results support the application of ML analysis in plant genomics and can help increase agricultural efficiency and production, leading to higher yields and more sustainable farming practices.

Modeling the mechanical properties of recycled aggregate concrete using hybrid machine learning algorithms

To explore the complicated functional relationship between key parameters such as the recycled aggregate properties, mix proportion and compressive strength of recycled aggregate concrete (RAC), a complete database involving 607 records from relevant published literature was built. Two standard algorithms (artificial neural network (ANN) and support vector regression (SVR)) and two optimized hybrid models (Particle Swarm Optimization based SVR (PSO-SVR) and grey Wolf optimizer based SVR (GWO-SVR)) were adopted. Furthermore, two interpretable algorithms (Partial Dependence Plot (PDP) and SHapley Additive exPlanations (SHAP)) were utilized to assess the global and local approaches independent of machine learning models, contributing towards decision-making rationales. Results indicated that the coefficient of determination (R2) of ANN, SVR, PSO-SVR and GWO-SVR were 0.7569, 0.5914, 0.8995 and 0.9056 respectively, showing that hybrid models outperformed the conventional models. However, GWO-SVR was the most problematic with overfitting when analyzing its three subsets. The two feature importance analyses revealed cement content, water content, natural fine aggregates, and water absorption as significant characteristics that affect mechanical performance.

On the Use of Low-Frequency Distributed Acoustic Sensing Data for In-Well Monitoring and Well Integrity: Qualitative Interpretation

Summary Distributed acoustic sensing (DAS) is used to monitor production and injection wells in real time. Here we present theory, data examples, and interpretation of low-frequency (LF)-DAS data acquired on fiber-optic (FO) cables clamped to the production tubing. We show that these data give unprecedented insights into the well and serve as a tool for qualitative analysis of well integrity, well operations, and production. By understanding these data, we show that we, for example, can detect leaks, monitor valve performance, and detect the liquid level in Annulus A. While in-well DAS’s sensitivity to changes in temperature over time is well known, its sensitivity to pressure and flow at low frequencies has not been well understood. Here, we show that the LF-DAS response is multifaceted and dependent on the fluid content of the different tubulars in a well, the casing, and completion design, and if a valve is open or closed. Very often a pressure response is seen due to an accompanied temperature effect; while this complicates the overall picture for quantitative data analysis, qualitative interpretation can be made simple. Understanding these processes in LF-DAS data forms the foundation for developing automated interpretation algorithms in the future.

Analyzing syntactic constructs of Java programs with machine learning

The massive number of open-source projects in public repositories has notably increased in the last years. Such repositories represent valuable information to be mined for different purposes, such as documenting recurrent syntactic constructs, analyzing the particular constructs used by experts and beginners, using them to teach programming and to detect bad programming practices, and building programming tools such as decompilers, Integrated Development Environments or Intelligent Tutoring Systems. An inherent problem of source code is that its syntactic information is represented with tree structures, while traditional machine learning algorithms use n-dimensional datasets. Therefore, we present a feature engineering process to translate tree structures into homogeneous and heterogeneous n-dimensional datasets to be mined. Then, we run different interpretable (supervised and unsupervised) machine learning algorithms to mine the syntactic information of more than 17 million syntactic constructs in Java code. The results reveal interesting information such as the Java constructs that are barely (and widely) used (e.g., bitwise operators, union types and static blocks), different language features and patterns mostly (and barely) used by beginners (and experts), the discovery of particular types of source code (e.g., helper or utility classes, data transfer objects and too complex abstractions), and how complexity is an inherent characteristic in some clusters of syntactic constructs.

Table interpretation of the temporal description logic LTLALC

Description logics are widely used to describe and represent knowledge in the Semantic Web. This is a modern and powerful mechanism that provides the possibility of extracting knowledge from already existing ones. Thanks to this, conceptual of subject areas modeling has become one of the fields of application of descriptive logics, taking into account the use of inference mechanisms. Conceptual modeling is used to create databases and knowledge bases. A key issue of the subject area modeling is the ability to monitor the dynamics of changes in the state of the subject area over time. It is necessary to describe not only the current actual state of the database (knowledge bases), but also the background. Temporal descriptive logics are used to solve this problem. They have the same set of algorithmic problems that are presented in conventional descriptive logics, but to them are added questions related to the description of knowledge in time. This refers to the form of time (continuous or discrete), time structure (moments of time, intervals, chains of intervals), time linearity (linear or branched), domain (present, past, future), the concept of “now”, the method of measurement, etc. An urgent task today is to create an algorithm for the temporal interpretation of conventional descriptive logics. That is, to show a way in which temporal descriptive logic can be applied to ordinary descriptive logic. The paper presents an algorithm for temporal interpretation of LTL into ALC. Linear, unbranched time is chosen for the description goal. It is presented in the form of a whole temporal axis with a given linear order on it. Only the future tense is considered. The algorithm contains graphic notations of LTL application in ALC: concepts, concept constructors, roles, role constructors, TBox and ABox. Numerous examples are used to illustrate the application of the algorithm.

Possibilities and Limitations of CNV Interpretation Software and Algorithms in Homo Sapiens

Background: Technical advances and cost reduction have allowed for the worldwide popularity of array platforms. Otherwise called “molecular karyotyping”, it yields a large amount of CNV data, which is useless without interpretation. Objective: This study aims to review existing CNV interpretation software and algorithms to reveal their possibilities and limitations. Results: Open and user-friendly CNV interpretation software is limited to several options, which mostly do not allow for cross-interpretation. Many algorithms are generally based on the Database of Genomic Variants, CNV size, inheritance data, and disease databases, which currently seem insufficient. Conclusion: The analysis of CNV interpretation software and algorithms resulted in a conclusion that it is necessary to expand the existing algorithms of CNV interpretation and at least include pathway and expression data. A user-friendly freely available CNV interpretation software, based on the expanded algorithms, is yet to be created.

Mortality prediction of elderly Asian patients with acute coronary syndrome (ACS) using interpretable machine learning algorithm

Mortality prediction of elderly Asian patients with acute coronary syndrome (ACS) using interpretable machine learning algorithm

Identifying potential candidates for advanced heart failure therapies using an interpretable machine learning algorithm.

Systems level barriers to heart failure (HF) care limit access to HF advanced therapies (heart transplantation, left ventricular assist devices). There is a need for automated systems that can help clinicians ensure patients with HF are evaluated for HF advanced therapies at the appropriate time to optimize outcomes. We performed a retrospective study using the REVIVAL (Registry Evaluation of Vital Information for VADs in Ambulatory Life) and INTERMACS (Interagency Registry for Mechanically Assisted Circulatory Support) registries. We developed a novel machine learning model based on principles of tropical geometry and fuzzy logic that can accommodate clinician knowledge and provide recommendations regarding need for advanced therapies evaluations that are accessible to end-users. The model was trained and validated using data from 4,694 HF patients. When initiated with clinical knowledge from HF and transplant cardiologists, the model achieved an F1 score of 43.8%, recall of 51.1%, and precision of 46.9%. The model achieved comparable performance compared with other commonly used machine learning models. Importantly, our model was 1 of only 3 models providing transparent and parsimonious clinical rules, significantly outperforming the other 2 models. Eleven clinical rules were extracted from the model which can be leveraged in clinical practice. A machine learning model capable of accepting clinical knowledge and making accessible recommendations was trained to identify patients with advanced HF. While this model was developed for HF care, the methodology has multiple potential uses in other important clinical applications.

Bayesian Rule Set: A Quantitative Alternative to Qualitative Comparative Analysis

We introduce the Bayesian Rule Set (BRS) as an alternative to Qualitative Comparative Analysis (QCA) when data are large and noisy. BRS is an interpretable machine learning algorithm that classifies observations using rule sets, which are conditions connected by logical operators (e.g., IF (condition A AND condition B) OR (condition C), THEN Y = TRUE). Like QCA, BRS is highly interpretable and capable of revealing complex nonlinear relationships in data. It also has several advantages over QCA: it is compatible with probabilistically generated data, it avoids overfitting and improves interpretability by making direct trade-offs between in-sample fitness and complexity, and it remains computationally efficient with many covariates. Our contributions are threefold: we modify the BRS algorithm to facilitate its usage in the social sciences, propose methods to quantify uncertainties of rule sets, and develop graphical tools for presenting rule sets. We illustrate these methods with two empirical examples from political science.

Interpretable Machine Learning Prediction of Drug-Induced QT Prolongation: Electronic Health Record Analysis.

Drug-induced long-QT syndrome (diLQTS) is a major concern among patients who are hospitalized, for whom prediction models capable of identifying individualized risk could be useful to guide monitoring. We have previously demonstrated the feasibility of machine learning to predict the risk of diLQTS, in which deep learning models provided superior accuracy for risk prediction, although these models were limited by a lack of interpretability. In this investigation, we sought to examine the potential trade-off between interpretability and predictive accuracy with the use of more complex models to identify patients at risk for diLQTS. We planned to compare a deep learning algorithm to predict diLQTS with a more interpretable algorithm based on cluster analysis that would allow medication- and subpopulation-specific evaluation of risk. We examined the risk of diLQTS among 35,639 inpatients treated between 2003 and 2018 with at least 1 of 39 medications associated with risk of diLQTS and who had an electrocardiogram in the system performed within 24 hours of medication administration. Predictors included over 22,000 diagnoses and medications at the time of medication administration, with cases of diLQTS defined as a corrected QT interval over 500 milliseconds after treatment with a culprit medication. The interpretable model was developed using cluster analysis (K=4 clusters), and risk was assessed for specific medications and classes of medications. The deep learning model was created using all predictors within a 6-layer neural network, based on previously identified hyperparameters. Among the medications, we found that class III antiarrhythmic medications were associated with increased risk across all clusters, and that in patients who are noncritically ill without cardiovascular disease, propofol was associated with increased risk, whereas ondansetron was associated with decreased risk. Compared with deep learning, the interpretable approach was less accurate (area under the receiver operating characteristic curve: 0.65 vs 0.78), with comparable calibration. In summary, we found that an interpretable modeling approach was less accurate, but more clinically applicable, than deep learning for the prediction of diLQTS. Future investigations should consider this trade-off in the development of methods for clinical prediction.

A Lightweight Model for Ship Detection and Recognition in Complex-Scene SAR Images

SAR ship detection and recognition are important components of the application of SAR data interpretation, allowing for the continuous, reliable, and efficient monitoring of maritime ship targets, in view of the present situation of SAR interpretation applications. On the one hand, because of the lack of high-quality datasets, most existing research on SAR ships is focused on target detection. Additionally, there have been few studies on integrated ship detection and recognition in complex SAR images. On the other hand, the development of deep learning technology promotes research on the SAR image intelligent interpretation algorithm to some extent. However, most existing algorithms only focus on target recognition performance and ignore the model’s size and computational efficiency. Aiming to solve the above problems, a lightweight model for ship detection and recognition in complex-scene SAR images is proposed in this paper. Firstly, in order to comprehensively improve the detection performance and deployment capability, this paper applies the YOLOv5-n lightweight model as the baseline algorithm. Secondly, we redesign and optimize the pyramid pooling structure to effectively enhance the target feature extraction efficiency and improve the algorithm’s operation speed. Meanwhile, to suppress the influence of complex background interference and ships’ distribution, we integrate different attention mechanism into the target feature extraction layer. In addition, to improve the detection and recognition performance of densely parallel ships, we optimize the structure of the model’s prediction layer by adding an angular classification module. Finally, we conducted extensive experiments on the newly released complex-scene SAR image ship detection and recognition dataset, named the SRSDDv1.0 dataset. The experimental results show that the minimum size of the model proposed in this paper is only 1.92 M parameters and 4.52 MB of model memory, which can achieve an excellent F1-Score performance of 61.26 and an FPS performance of 68.02 on the SRSDDv1.0 dataset.

Statistical Analysis of the Potential of Landslides Induced by Combination between Rainfall and Earthquakes

This study analyzed the potential of landslides induced by the interaction between rainfall and earthquakes. Dapu Township and Alishan Township in Chiayi County, southern Taiwan, were included as study areas. From satellite images and the literature, we collected data for multiple years and time series and then used the random forest data mining algorithm for satellite image interpretation. A hazard index for the interaction between earthquakes and rainfall (IHERI) was proposed, and an index for the degree of land disturbance (IDLD) was estimated to explore the characteristics of IHERI under specific natural environmental and slope land use conditions. The results revealed that among the investigated disaster-causing factors, the degree of slope land use disturbance, the slope of the natural environment, and rainfall exerted the strongest effect on landslide occurrence. When IHERI or IDLD was higher, the probability of a landslide also increased, and under conditions of a similar IDLD, the probability of landslides increased as the IHERI value increased, and vice versa. Thus, given the interaction between rainfall and earthquakes in the study area, the effect of the degree of slope land use disturbance on landslides should not be ignored. The results of a receiver operating characteristic (ROC) curve analysis indicated that the areas under the ROC curve for landslides induced by different trigger factors were all above 0.94. The results indicate that the area in which medium–high-level landslides are induced by an interaction between rainfall and earthquakes is large.

RADT-15. 18F-DOPA PET/CT SURVEILLANCE FOR GLIOBLASTOMA: A RADIOMIC MODEL FROM A PROSPECTIVE PHASE II CLINICAL TRIAL PREDICTING SURVIVAL IN IDH-WILDTYPE, MGMT-UNMETHYLATED PATIENTS

Abstract BACKGROUND Interpretation of serial magnetic resonance imaging (MRI) for glioblastoma following radiation therapy (RT) is complicated by difficulty differentiating tumor from treatment-related changes, even using updated RANO criteria. The incorporation of novel imaging such as 3,4-dihydroxy-6-[18F]-fluoro-L-phenylalanine (18F-DOPA) PET/CT to post-treatment serial imaging may improve prognostication and better facilitate future treatment decisions. METHODS The secondary analysis of a recent phase II prospective clinical trial of 18F-DOPA PET/CT-directed dose-escalated RT included patients with IDH-wildtype, MGMT-unmethylated glioblastoma who underwent post-treatment serial 18F-DOPA PET/CT surveillance. Quantitative features were extracted from pre-RT and post-RT serial PET/CT images, and robust prognostic features were selected using an in-house workflow. Both an automated machine learning (ML) algorithm and an interpretable ML algorithm were utilized to correlate surveillance PET image features with subsequent survival of greater than 12 months versus less than 12 months from the surveillance timepoint. Changes from pre-RT to post-RT PET/CT (delta model) were also assessed for association with post-RT survival and validated with a separate cohort. RESULTS Thirty-five patients with IDH-wildtype, MGMT-unmethylated glioblastoma who had at least one available (range: 1-14) post-treatment 18F-DOPA PET/CT were included. Twenty-four were used for model training, while 11 were used for validation. Ultimately, a five-feature post-RT model utilizing two shape, two texture, and one first-order radiomic feature was selected. For the delta model, five texture, two shape, and one first order radiomic feature were selected. The models show 90% accuracy in predicting survival < 12 months post-surveillance on the training set, and 68%-73% accuracy (AUC 0.64-0.73) for the validation cohort. Delta features were significantly associated with overall survival (p < 0.05). CONCLUSIONS Post-RT serial 18F-DOPA PET/CT imaging can help distinguish true tumor progression in patients with glioblastoma using a radiomics model. Tumor response evaluated for changes from pre-RT to post-RT 18F-DOPA PET/CT also predicted subsequent overall survival.

Essential insights into decision mechanism of landslide susceptibility mapping based on different machine learning models

This work aims to discuss and compare the inherent essence of different machine learning algorithms for landslide susceptibility models (LSMs), which is of great significance for accurate prevention and detection of landslides. A geospatial database was established in GIS based on various factors of topography, geological conditions, environmental conditions and human activities, including 22 conditioning factors and 866 historical landslides. As for model algorithms, ANN is an operation model composed of a large number of interconnected nodes, and RF refers to an ensemble method of separately trained binary decision trees. Two algorithms were adopted in this paper for landslide susceptibility models. Meantime, an interpretable algorithm SHAP was used to gain insight into essential decision mechanism of LSMs. The result showed that RF model exhibits better stability and robustness. The global interpretation shows that the same landslide moderators play different roles in different models. The local interpretation shows that for the same evaluation unit different models give different decision mechanisms, and the local interpretation can be combined with the field survey, which can provide a comprehensive framework for assessing the assigned landslide.

Abstract 13558: Opening the Black Box of Artificial Intelligence: Visualization of Detecting Heart Failure Subtypes Using Electrocardiography

Objectives: To evaluate and visualize how artificial intelligence (AI) recognizes electrocardiography (ECG) characteristics of each heart failure (HF) subtype. We developed a visually explainable and interpretable AI-ECG algorithm for various HF subtypes to meet this need. This study proposes the ShapeExplainer (SE) to interpret models. Methods: From 33,920 patients at two hospitals, we trained a convolutional neural network to identify patients with HF subtypes using ECG and clinical diagnosis. HF subtypes include heart failure with reduced ejection fraction (HFrEF), heart failure with mid-range ejection fraction (HFmrEF, and heart failure with preserved ejection fraction (HFpEF) ( Figure a ). To interprete each HF classifier model, SE, which converts non-HF group data to be recognized as each HF subtype by each classifier, was trained. The converted fake ECG shows what the change is for the classifier to recognize as each HF subtype. SE from the non-HF group of the external validation set shows the result by comparing the shape before and after transformation. Results: When tested on an independent set of 6,784 patients, the network model yielded values for the area under the curves of HFrEF of 0.945, HFmrEF of 0.884, and HFpEF of 0.859, respectively. The results of each classifier were visualized by overlapping the original and newly generated ECGs. Comparing each subtype and its fake ECG, there was no clear trend of electrocardiographic features. However, fake ECG showed a tendency of qrs amplitude as it progressed to non-HF, HFpEF, HFmrEF, and HFrEF, especially in T wave or ST-change ( Figure b ). Conclusions: The proposed fake ECG generating method is a new approach for interpreting the AI-ECG model, helping physicians understand how AI recognizes HF subtypes.

12. Expanding clinical actionability in individual patient profiles with the Molecular Oncology Almanac

<h3>Background</h3> MOAlmanac is an open source clinical interpretation algorithm and paired knowledge base for precision cancer medicine used to rapidly characterize and identify genomic features related to therapeutic sensitivity and resistance and of prognostic relevance. <h3>Methods</h3> MOAlmanac assesses tumor actionability by evaluating individual genomic features (e.g. somatic or germline variants, copy number alterations, or fusions), interactions between them, and secondary features (such as mutational burden, mutational signatures, and MSI status) relative to our knowledge base. In addition, it also performs profile-to-cell line matchmaking to nominate genomically similar cell lines for hypothesis expansion. All clinically relevant findings are summarized into a web-based report. The underlying knowledge base can be accessed through our API endpoints and web browser, and we developed a cloud-based web portal to increase accessibility of the algorithm. <h3>Results</h3> 32,108 samples from 66 cancer types received targeted sequencing with Oncopanel and were evaluated by MOAlmanac. Using Oncopanel's tier 1 and tier 2 criteria for clinical actionability, we observed that 8,285 samples harbored at least one alteration suggesting therapeutic sensitivity based on FDA approvals or clinical guidelines, increasing to 18,117 when considering clinical trials, clinical, preclinical, and inferential evidence. <h3>Conclusion</h3> Clinical actionability of molecular tumor data was increased in individual patients by expanding the set of evidence considered. Source code and a web portal for this project are available at moalmanac.org.

Biomedical Applications using Hand Gesture with Electromyography Control Signal

Wearables developed for human body signal detection receive increasing attention in the current decade. Compared to implantable sensors, wearables are more focused on body motion detection, which can support human–machine interaction (HMI) and biomedical applications. In wearables, electromyography (EMG), force myography (FMG), and electrical impedance tomography (EIT) based body information monitoring technologies are broadly presented. In the literature, all of them have been adopted for many similar application scenarios, which easily confuses researchers when they start to explore the area. Hence, in this article, we review the three technologies in detail, from basics including working principles, device architectures, interpretation algorithms, application examples, merits and drawbacks, to state-of-the-art works, challenges remaining to be solved and the outlook of the field. We believe the content in this paper could help readers create a whole image of designing and applying the three technologies in relevant scenarios. Index Terms : FMG; EMG; EIT; biological signal; human–system interactivities.

30 days mortality prediction and risk factor analysis of Asian patients with ACS using interpretable machine learning algorithm

Abstract Background Thrombolysis in Myocardial Infarction (TIMI) is used to predict the mortality rate in patients with acute coronary syndrome (ACS). TIMI was developed with limited data on the Asian cohort and was based on the Western cohort. STEMI and NSTEMI have separate TIMI scores. There has been limited research on Asian ACS patients using interpretable machine learning (ML) algorithms. Purpose To construct a single 30-day mortality risk scoring system, as well as identify and analyse risk factors in ASIAN patients with ACS, that is applicable to both STEMI and NSTEMI patients, using an interpretable ML algorithm. Methods The National Cardiovascular Disease Database registry data of 9054 patients was used. 70% of the data was used for algorithm development, with the remaining 30% used for validation Fifty-four parameters were considered, demographics, cardiovascular risk, medications, and clinical variables. To provide better guidance and advice for clinical judgement, the gradient boosting algorithm (XGBoost) for classification analysis and SHapley Additive exPlanation (SHAP) value analysis graphs were used. Each indicator's SHAP value indicates the impact on model output (mortality) and was calculated using the XGBoost model. The performance evaluation metric was the area under the curve (AUC). The model was validated with a validation dataset and compared to the conventional score TIMI for STEMI and NSTEMI. Results The performance on validation dataset of the XGBoost algorithm using the top ten predictors from SHAP for; STEMI (AUC = 0.8534, 95% CI: 0.8226–0.8842, Accuracy: 0.8053, Sensitivity: 0.73125, Specificity: 0.81355) and NSTEMI (AUC = 0.8145, 95% CI: 0.77–0.8589, Accuracy: 0.7972, Sensitivity: 0.64356, Specificity: 0.81232) outperformed TIMI score (STEMI AUC = 0.785, NSTEMI AUC = 0.543). Killip class, age, heart rate, fasting blood glucose, ACEI, creatine kinase, systolic blood pressure, HDLC, cardiac catheterization, and oralhypogly are the top ten predictors chosen by the SHAP feature selection in ascending order. Cardiac catheterization and pharmacotherapy drugs as selected predictors improve mortality prediction in STEMI and NSTEMI patients compared to TIMI. The variable names are displayed on the y-axis in ascending order of importance. The average SHAP value is shown next to them. The SHAP value is shown on the x-axis. The colour represents the value of the feature, ranging from small to large, allowing comprehension of the distribution of the SHAP values for each feature (Figure 1). We can see that having a high killip class and being older are linked to a lower survival rate in ACS patients. Cardiac catheterization procedures, as well as the use of ACEI and OHA, both improve patient mortality (Figure 2). Conclusions A single algorithm would classify ACS patients better than TIMI, which requires two distinct scores. In order to better predict 30-day mortality in an ASIAN population, interpretable ML can be used. Funding Acknowledgement Type of funding sources: Public grant(s) – National budget only. Main funding source(s): Technology Development Fund 1

Inside the “brain” of an artificial neural network: an interpretable deep learning approach to paroxysmal atrial fibrillation diagnosis from electrocardiogram signals during sinus rhythm

Abstract Background With the ongoing, rapid advances in Deep Learning (DL), such solutions can now detect medical conditions even invisible to the human eye. In this direction, efforts have been made to develop DL algorithms that diagnose paroxysmal atrial fibrillation (PAF) from electrocardiogram (ECG) signals in sinus rhythm (SR). However, many of the available approaches function as “black boxes”, with physicians unable to understand and trust their predictions. Purpose To train a DL model to detect PAF patients while in SR and apply an algorithm that interprets and visualises its decisions. Methods We obtained ECG samples from PAF and non-PAF patients during SR, from the PAF Prediction Challenge Database. After discarding unannotated samples and augmenting the sample size (by dividing each signal into 30-second segments), we split the whole dataset into a train (68%), a validation (16%) and a test (16%) set. No pair of samples belonging to different sets originated from the same patient. We trained the InceptionTime neural network on the train/validation sets and tested on the “unseen” test set after “hiding” the correct answers. Its performance was evaluated with the following metrics: Accuracy, f1-score, precision and recall (sensitivity). After repeating this process 20 times, we obtained a distribution for each score. Finally, we adjusted the Grad-CAM interpretation algorithm to our data and used it to visualise the areas perceived as important by the model. Results After pre-processing, 4,080, 30-second, two-lead ECG signals were allocated to the train set, 960 to the validation and 960 to the test set. Each subset contained an equal number of PAF and non-PAF samples. After repeated training and testing, we obtained a median accuracy of 0.84 (interquartile range, IQR: 0.66–0.88), an f1-score of 0.82 (IQR: 0.68–0.88) and a median precision and recall equal to 0.93 (IQR: 0.67–0.99) and 0.77 (IQR: 0.68–0.93), respectively. The Grad-CAM technique highlighted the ECG areas of interest that led to each decision. We selected and present both PAF-positive and -negative samples, perceived either correctly or falsely. Interestingly, correct model decisions tend to focus on the P-wave, while false ones fixate on other regions. Conclusions Although a pilot study with considerable limitations (small sample size, disregard of possible confounding due to comorbidities or other factors), this work shows how DL can be employed to distinguish between PAF and non-PAF patients from SR ECG samples, and confirms the potential of DL-enabled approaches to offer novel diagnostic capabilities. Most importantly, our effort provides a comprehensible, visual interpretation of the model's decisions. Demystifying DL behaviour can, not only improve such efforts by explaining false decisions, but also cultivate trust among clinicians and, possibly, point out directions for future research, since we can now see through the magnifying lens of a neural network. Funding Acknowledgement Type of funding sources: None.

Improving the accuracy of clinical interpretation of serological testing for the diagnosis of acute hepatitis a infection.

Hepatitis A (HA) remains a common infection globally that results in a self-limiting hepatitis with gastrointestinal and systemic symptoms. Direct detection of the virus in blood or stool is often not possible once symptomatic. Serological testing is frequently performed to investigate abnormal liver function tests - and interpretation of equivocal and low-level positive anti-hepatitis A virus (HAV) IgM is difficult even in the context of accurate epidemiological and clinical information. The aim of this project was to characterise the association between low-level reactive anti-HAV IgM results and clinical disease. Anti-HAV serology results recorded over 22 months were analysed. Equivocal and positive Architect anti-HAV IgM results were matched with available clinical and demographical data to identify confirmed and probable cases of acute HAV infection. Reactive anti-HAV IgM results were recorded for 88/7661 (1.15%) samples. Using clinical and laboratory data, 35 patients were confirmed to have acute HAV infection. Acute HA was associated with a mean Architect anti-HAV IgM value of 9.4 (SD 6.8-12.0). Cases of HA had a mean peak ALT value of 1920 (SD 682-3158). All confirmed cases (35/35) of acute HAV were associated with at least one clinical indicator, with 28/31 cases (90%) having a documented jaundice. All 35 (100%) cases of acute HAV infection had anti-HAV IgM > 4.0. A diagnosis other than acute HA was identified in 7/11 (63.6%) of low-level reactive anti-HAV IgM results (clinical data unavailable for further 4/11, 36.3%). Where clinical information was available, acute HA was excluded in all 31 patients with equivocal or low-level reactive anti-HAV IgM results. The accuracy of reports sent out to the clinician showed room for improvement. An interpretive algorithm is proposed including a clinically significant cut-off value for anti-HAV IgM.