Input Feature Set Research Articles

Optimizing the input feature sets and machine learning algorithms for reliable and accurate estimation of continuous, cuffless blood pressure

The advent of mobile devices, wearables and digital healthcare has unleashed a demand for accurate, reliable, and non-interventional ways to measure continuous blood pressure (BP). Many consumer products claim to measure BP with a cuffless device, but their lack of accuracy and reliability limit clinical adoption. Here, we demonstrate how multimodal feature datasets, comprising: (i) pulse arrival time (PAT); (ii) pulse wave morphology (PWM), and (iii) demographic data, can be combined with optimized Machine Learning (ML) algorithms to estimate Systolic BP (SBP), Diastolic BP (DBP) and Mean Arterial Pressure (MAP) within a 5 mmHg bias of the gold standard Intra-Arterial BP, well within the acceptable limits of the IEC/ANSI 80601-2-30 (2018) standard. Furthermore, DBP’s calculated using 126 datasets collected from 31 hemodynamically compromised patients had a standard deviation within 8 mmHg, while SBP’s and MAP’s exceeded these limits. Using ANOVA and Levene’s test for error means and standard deviations, we found significant differences in the various ML algorithms but found no significant differences amongst the multimodal feature datasets. Optimized ML algorithms and key multimodal features obtained from larger real-world data (RWD) sets could enable more reliable and accurate estimation of continuous BP in cuffless devices, accelerating wider clinical adoption.

Bombardier Aftermarket Demand Forecast with Machine Learning

Intermittent demand patterns are commonly present in business aircraft spare parts supply chains. Because of the infrequent arrivals and large variations in demand, aircraft aftermarket demand is difficult to forecast, which often leads to shortages or overstocking of spare parts. In this paper, we present the development and implementation of an advanced analytics framework at Bombardier Aerospace, which is carried out by the Bombardier inventory planning team and IVADO Labs to improve the aftermarket demand forecasting process. This integrated predictive analytics pipeline leverages machine-learning (ML) models and traditional time series models in a single framework in a systematic fashion. We also make use of a tree-based machine-learning method with a large set of input features to estimate two components of intermittent demand, namely demand sizes and interdemand intervals. Through the ML models, we incorporate different features, including those derived from flight data. Outputs of different forecasting models are combined using an ensemble technique that enhances the robustness and accuracy of the forecasts for different groups of aftermarket spare parts categorized by demand patterns. The validation results show an improvement in forecast accuracy of approximately 7% and in unbiased forecast of 5%. The ML-based Bombardier Aftermarket forecasting system has been successfully deployed and used to forecast the aftermarket demand at Bombardier of more than 1 billion Canadian dollars on a regular basis. History: This paper was refereed.

An explainable deep learning-enabled intrusion detection framework in IoT networks

Although the field of eXplainable Artificial Intelligence (XAI) has a significant interest these days, its implementation within cyber security applications still needs further investigation to understand its effectiveness in discovering attack surfaces and vectors. In cyber defence, especially anomaly-based Intrusion Detection Systems (IDS), the emerging applications of machine/deep learning models require the interpretation of the models' architecture and the explanation of models' prediction to examine how cyberattacks would occur. This paper proposes a novel explainable intrusion detection framework in the Internet of Things (IoT) networks. We have developed an IDS using a Short-Term Long Memory (LSTM) model to identify cyberattacks and explain the model's decisions. This uses a novel set of input features extracted by a novel SPIP (S: Shapley Additive exPlanations, P: Permutation Feature Importance, I: Individual Conditional Expectation, P: Partial Dependence Plot) framework to train and evaluate the LSTM model. The framework was validated using the NSL-KDD, UNSW-NB15 and TON_IoT datasets. The SPIP framework achieved high detection accuracy, processing time, and high interpretability of data features and model outputs compared with other peer techniques. The proposed framework has the potential to assist administrators and decision-makers in understanding complex attack behaviour.

Visualizing the Residue Interaction Landscape of Proteins by Temporal Network Embedding.

Characterizing the structural dynamics of proteins with heterogeneous conformational landscapes is crucial to understanding complex biomolecular processes. To this end, dimensionality reduction algorithms are used to produce low-dimensional embeddings of the high-dimensional conformational phase space. However, identifying a compact and informative set of input features for the embedding remains an ongoing challenge. Here, we propose to harness the power of Residue Interaction Networks (RINs) and their centrality measures, established tools to provide a graph theoretical view on molecular structure. Specifically, we combine the closeness centrality, which captures global features of the protein conformation at residue-wise resolution, with EncoderMap, a hybrid neural-network autoencoder/multidimensional-scaling like dimensionality reduction algorithm. We find that the resulting low-dimensional embedding is a meaningful visualization of the residue interaction landscape that resolves structural details of the protein behavior while retaining global interpretability. This feature-based graph embedding of temporal protein graphs makes it possible to apply the general descriptive power of RIN formalisms to the analysis of protein simulations of complex processes such as protein folding and multidomain interactions requiring no protein-specific input. We demonstrate this on simulations of the fast folding protein Trp-Cage and the multidomain signaling protein FAT10. Due to its generality and modularity, the presented approach can easily be transferred to other protein systems.

Chess Position Evaluation Using Radial Basis Function Neural Networks

The game of chess is the most widely examined game in the field of artificial intelligence and machine learning. In this work, we propose a new method for obtaining the evaluation of a chess position without using tree search and examining each candidate move separately, like a chess engine does. Instead of exploring the search tree in order to look several moves ahead, we propose to use the much faster and less computationally demanding estimations of a properly trained neural network. Such an approach offers the benefit of having an estimation for the position evaluation in a matter of milliseconds, while the time needed by a chess engine may be several orders of magnitude longer. The proposed approach introduces models based on the radial basis function (RBF) neural network architecture trained with the fuzzy means algorithm, in conjunction with a novel set of input features; different methods of network training are also examined and compared, involving the multilayer perceptron (MLP) and convolutional neural network (CNN) architectures and a different set of input features. All methods were based upon a new dataset, which was developed in the context of this work, derived by a collection of over 1500 top-level chess games. A Java application was developed for processing the games and extracting certain features from the arising positions in order to construct the dataset, which contained data from over 80,000 positions. Various networks were trained and tested as we considered different variations of each method regarding input variable configurations and dataset filtering. Ultimately, the results indicated that the proposed approach was the best in performance. The models produced with the proposed approach are suitable for integration in model-based decision-making frameworks, e.g., model predictive control (MPC) schemes, which could form the basis for a fully-fledged chess-playing software.

Multivariate Time Series Predictor With Parameter Optimization and Feature Selection Based on Modified Binary Salp Swarm Algorithm

More and more time series data appear in various fields, and the prediction of multivariate time series has been the key to solve many industrial problems. Therefore, it is necessary to establish an accurate prediction model. As an efficient recursive neural network, an echo state network (ESN) model has been widely used in time series prediction. However, it usually faces the problem of how to choose suitable reservoir parameters for different applications. In addition, selecting the input feature set is also an important issue, which will affect the accuracy and computational efficiency of the prediction model. To solve these problems, the modified binary salp swarm algorithm-based optimization ESN (MBSSA-ESN) is proposed for multivariate time series prediction, which can simultaneously realize feature subset selection and parameter optimization. In order to verify the effectiveness of the proposed model, Beijing air quality index data are used for simulation and the key index PM2.5 is used as the target variable for experiment. Compared with several related methods, the proposed model achieves the best results in all evaluation indicators, indicating that the MBSSA-ESN model is competitive in the task of multivariate time series prediction.

Cross-model consensus of explanations and beyond for image classification models: an empirical study

Existing interpretation algorithms have found that, even deep models make the same and right predictions on the same image, they might rely on different sets of input features for classification. However, among these sets of features, some common features might be used by the majority of models. In this paper, we are wondering what are the common features used by various models for classification and whether the models with better performance may favor those common features. For this purpose, our works uses an interpretation algorithm to attribute the importance of features (e.g., pixels or superpixels) as explanations, and proposes the cross-model consensus of explanations to capture the common features. Specifically, we first prepare a set of deep models as a committee, then deduce the explanation for every model, and obtain the consensus of explanations across the entire committee through voting. With the cross-model consensus of explanations, we conduct extensive experiments using 80+ models on 5 datasets/tasks. We find three interesting phenomena as follows: (1) the consensus obtained from image classification models is aligned with the ground truth of semantic segmentation; (2) we measure the similarity of the explanation result of each model in the committee to the consensus (namely consensus score), and find positive correlations between the consensus score and model performance; and (3) the consensus score coincidentally correlates to the interpretability.

Robust Neural Particle Identification Models

The volume of data processed by the Large Hadron Collider experiments demands sophisticated selection rules typically based on machine learning algorithms. One of the shortcomings of these approaches is their profound sensitivity to the biases in training samples. In the case of particle identification (PID), this might lead to degradation of the efficiency for some decays not present in the training dataset due to differences in input kinematic distributions. In this talk, we propose a method based on the Common Specific Decomposition that takes into account individual decays and possible misshapes in the training data by disentangling common and decay specific components of the input feature set. We show that the proposed approach reduces the rate of efficiency degradation for the PID algorithms for the decays reconstructed in the LHCb detector.

Predicting Patient Satisfaction With Medications for Treating Opioid Use Disorder: Case Study Applying Natural Language Processing to Reviews of Methadone and Buprenorphine/Naloxone on Health-Related Social Media.

Medication-assisted treatment (MAT) is an effective method for treating opioid use disorder (OUD), which combines behavioral therapies with one of three Food and Drug Administration-approved medications: methadone, buprenorphine, and naloxone. While MAT has been shown to be effective initially, there is a need for more information from the patient perspective about the satisfaction with medications. Existing research focuses on patient satisfaction with the entirety of the treatment, making it difficult to determine the unique role of medication and overlooking the views of those who may lack access to treatment due to being uninsured or concerns over stigma. Studies focusing on patients' perspectives are also limited by the lack of scales that can efficiently collect self-reports across domains of concerns. A broad survey of patients' viewpoints can be obtained through social media and drug review forums, which are then assessed using automated methods to discover factors associated with medication satisfaction. Because the text is unstructured, it may contain a mix of formal and informal language. The primary aim of this study was to use natural language processing methods on text posted on health-related social media to detect patients' satisfaction with two well-studied OUD medications: methadone and buprenorphine/naloxone. We collected 4353 patient reviews of methadone and buprenorphine/naloxone from 2008 to 2021 posted on WebMD and Drugs.com. To build our predictive models for detecting patient satisfaction, we first employed different analyses to build four input feature sets using the vectorized text, topic models, duration of treatment, and biomedical concepts by applying MetaMap. We then developed six prediction models: logistic regression, Elastic Net, least absolute shrinkage and selection operator, random forest classifier, Ridge classifier, and extreme gradient boosting to predict patients' satisfaction. Lastly, we compared the prediction models' performance over different feature sets. Topics discovered included oral sensation, side effects, insurance, and doctor visits. Biomedical concepts included symptoms, drugs, and illnesses. The F-score of the predictive models across all methods ranged from 89.9% to 90.8%. The Ridge classifier model, a regression-based method, outperformed the other models. Assessment of patients' satisfaction with opioid dependency treatment medication can be predicted using automated text analysis. Adding biomedical concepts such as symptoms, drug name, and illness, along with the duration of treatment and topic models, had the most benefits for improving the prediction performance of the Elastic Net model compared to other models. Some of the factors associated with patient satisfaction overlap with domains covered in medication satisfaction scales (eg, side effects) and qualitative patient reports (eg, doctors' visits), while others (insurance) are overlooked, thereby underscoring the value added from processing text on online health forums to better understand patient adherence.

Cotton Fiber Quality Estimation Based on Machine Learning Using Time Series UAV Remote Sensing Data

As an important factor determining the competitiveness of raw cotton, cotton fiber quality has received more and more attention. The results of traditional detection methods are accurate, but the sampling cost is high and has a hysteresis, which makes it difficult to measure cotton fiber quality parameters in real time and at a large scale. The purpose of this study is to use time-series UAV (Unmanned Aerial Vehicle) multispectral and RGB remote sensing images combined with machine learning to model four main quality indicators of cotton fibers. A deep learning algorithm is used to identify and extract cotton boll pixels in remote sensing images and improve the accuracy of quantitative extraction of spectral features. In order to simplify the input parameters of the model, the stepwise sensitivity analysis method is used to eliminate redundant variables and obtain the optimal input feature set. The results of this study show that the R2 of the prediction model established by a neural network is improved by 29.67% compared with the model established by linear regression. When the spectral index is calculated after removing the soil pixels used for prediction, R2 is improved by 4.01% compared with the ordinary method. The prediction model can well predict the average length, uniformity index, and micronaire value of the upper half. R2 is 0.8250, 0.8014, and 0.7722, respectively. This study provides a method to predict the cotton fiber quality in a large area without manual sampling, which provides a new idea for variety breeding and commercial decision-making in the cotton industry.

Explainable Rules and Heuristics in AI Algorithm Recommendation Approaches—A Systematic Literature Review and Mapping Study

The exponential use of artificial intelligence (AI) to solve and automated complex tasks has catapulted its popularity generating some challenges that need to be addressed. While AI is a powerful means to discover interesting patterns and obtain predictive models, the use of these algorithms comes with a great responsibility, as an incomplete or unbalanced set of training data or an unproper interpretation of the models’ outcomes could result in misleading conclusions that ultimately could become very dangerous. For these reasons, it is important to rely on expert knowledge when applying these methods. However, not every user can count on this specific expertise; non-AI-expert users could also benefit from applying these powerful algorithms to their domain problems, but they need basic guidelines to obtain the most out of AI models. The goal of this work is to present a systematic review of the literature to analyze studies whose outcomes are explainable rules and heuristics to select suitable AI algorithms given a set of input features. The systematic review follows the methodology proposed by Kitchenham and other authors in the field of software engineering. As a result, 9 papers that tackle AI algorithm recommendation through tangible and traceable rules and heuristics were collected. The reduced number of retrieved papers suggests a lack of reporting explicit rules and heuristics when testing the suitability and performance of AI algorithms.

Recognition of human emotion transition from video sequence using triangulation induced various centre pairs distance signatures

Human emotion recognition manifests the concept of contextual understanding of human communication in nonverbal form. This fascinates us to monitor the dynamic information of an emotion. In this article, we present an automatic recognition system for observing gradual changes in human emotion available in the form of a video sequence. In our proposed system, we introduce a triangulation mechanism for geometric feature generation from salient landmarks on the face image. Here, well known active appearance model (AAM) is applied to consecutive image frames in a sequence to track all locations of landmark points. In this work, we explore four centre points (Incentre, Circumcentre, Barycentre, and Orthocentre) from all the triangles formed by salient facial points selected in a triplet. Six different geometric distance signatures viz. Incentre–Circumcentre, Incentre–Barycentre, Incentre–Orthocentre, Circumcentre–Barycentre, Circumcentre–Orthocentre, and Barycentre–Orthocentre are extracted accordingly by taking all possible pairs of those four centres into account to signify several ways of the geometric representation of an emotional image sequence. The discriminative merit of each distance signature is analysed independently by employing them separately as input feature sets into a multilayer perceptron (MLP) classifier for recognition of six basic emotions(anger, disgust, fear, happiness, sadness, and surprise). We verify the performance of our proposed system through the experimentation on several publically available benchmark video datasets such as Extended Cohn–Kanade (CK＋), M&M Initiative (MMI), and Multimedia Understanding Group (MUG). Experimental results indicate impressive accuracy of 98.47% for CK＋, 92.57% for MMI, and 98.62% for the MUG dataset on Incentre–Circumcentre pair distance signature. Moreover, to justify the effectiveness of our proposed method we compare our experimental results with other existing states of the art available in the literature.

On the Performance of Machine Learning Based Flight Delay Prediction – Investigating the Impact of Short-Term Features

People and companies today are connected around the world, which has led to a growing importance of the aviation industry. As flight delays are a big challenge in aviation, machine learning algorithms can be used to forecast those. This paper investigates the prediction of the occurrence of flight arrival delays with three prominent machine learning algorithms for a data set of domestic flights in the USA. The task is regarded as a classification problem. The focus lies on the investigation of the influence of short-term features on the quality of the results. Therefore, three scenarios are created that are characterised by different input feature sets. When forgoing the inclusion of short-term information in order to shift the prediction timing to an early point in time, an accuracy of 69.5% with a recall of 68.2% is achieved. By including information on the delay that the aircraft had on its previous flight, the prediction quality increases slightly. Hence, this is a compromise between the early prediction timing of the first model and the good prediction quality of the third model, where the departure delay of the aircraft is added as an input feature. In this case, an accuracy of 89.9% with a recall of 83.4% is obtained. The desired timing of prediction therefore determines which features to use as inputs since short-term features significantly improve the prediction quality.

Adaptive and Lightweight Network Traffic Classification for Edge Devices

Network traffic classification (NTC) is essential to network measurement, management, and security. Recently, Artificial Intelligence (AI)-based NTCs have demonstrated its accuracy in identifying data traffic. However, existing AI-based NTCs require inputs of lengthy data payload with padding, which can lead to complex designs of NTC models that are especially challenging to networking edge devices, e.g., Wi-Fi routers, etc. To address this issue, an adaptive and lightweight NTC framework is proposed in this paper. In specific, an input feature contribution extraction scheme is developed to weigh each input feature based on both the significance and the uniqueness of the corresponding feature. The optimal set of input features are determined to minimize the complexity of a targeting AI-based NTC while maintaining high performance in classification. Moreover, an autonomous update scheme is proposed to detect the changes in feature contribution and process updates. Evaluations on two fundamental AI-based NTC models, i.e., MLP and CNN, demonstrated that the proposed scheme can significantly reduce the input features and accelerate NTC models by one to two magnitudes while maintaining high accuracy. The proposed autonomous update scheme can accurately detect a change in feature contributions and update the NTC models to sustain the high accuracy.

Hierarchical, multiclass prediction of Alzheimer’s clinical diagnosis using imputed, multimodal NACC data

AbstractBackgroundEarly detection and diagnosis are critical for effective treatment of Alzheimer’s disease (AD). Statistical machine learning holds promise for disease prediction using large AD repositories, such as the National Alzheimer’s Coordinating Center (NACC) dataset. Past work has extensively explored binary classifications of AD diagnosis. Here, we extended this to address two major, remaining challenges: data imputation and multiclass classification (e.g., AD vs Mild Cognitive Impairment, MCI vs. Healthy Cognition, HC).MethodFor imputation, we tested single‐value (baseline), k‐NN, and Bayesian imputation on a subset of complete training data with artificially removed values (completely at random). Per modality of interest, a systematic search was conducted for the best input feature set to impute the missing values. The mean absolute deviation and mismatch rate between the imputed and true values were used to evaluate the methods on a held‐out test set. For classification, we then implemented both “flat” multiclass and hierarchical classification, an extension of multiclass classification that organizes classes into a tree, to predict clinical diagnosis using multimodal inputs. Several pipelines with combinations of imputation methods, classifiers, and input modalities were tested on different hierarchical classification strategies, e.g., HC vs [AD vs MCI], compared with the flat, multiclass baseline (AD vs MCI vs HC).ResultCompared to the baseline methods for both imputation and classification, the proposed alternatives performed better, as measured by the predetermined imputation and classification metrics evaluated on the test set. For imputation, behavioral and neuropsychiatric features were more easily and accurately imputed while ApoE genotype was more difficult to impute. For classification, using a combination of classifiers and input features in the hierarchy outperformed the baseline method and support the notion that select biomarkers are better able to characterize AD subpopulations within the hierarchy.ConclusionEstablishing a systematic imputation method tailored to AD biomarker modalities fills a critical gap in achieving more robust disease prediction. Building a hierarchical classifier parallels working through a differential diagnosis for AD, which offers greater translational value than binary classification of AD. In combination, these results form an important step toward developing reliable, holistic classifiers of AD disease status.

A Framework for Auditing Multilevel Models using Explainability Methods

Multilevel models (MLMs) are increasingly deployed in industry across different functions. Applications usually result in binary classification within groups or hierarchies based on a set of input features. For transparent and ethical applications of such models, sound audit frameworks need to be developed. In this paper, an audit framework for technical assessment of regression MLMs is proposed. The focus is on three aspects: model, discrimination, and transparency & explainability. These aspects are subsequently divided into sub-aspects. Contributors, such as inter MLM-group fairness, feature contribution order, and aggregated feature contribution, are identified for each of these sub-aspects. To measure the performance of the contributors, the framework proposes a shortlist of KPIs, among others, intergroup individual fairness (DiffInd_MLM) across MLM-groups, probability unexplained (PUX) and percentage of incorrect feature signs (POIFS). A traffic light risk assessment method is furthermore coupled to these KPIs. For assessing transparency & explainability, different explainability methods (SHAP and LIME) are used, which are compared with a model intrinsic method using quantitative methods and machine learning modelling.
 Using an open-source dataset, a model is trained and tested and the KPIs are computed. It is demonstrated that popular explainability methods, such as SHAP and LIME, underperform in accuracy when interpreting these models. They fail to predict the order of feature importance, the magnitudes, and occasionally even the nature of the feature contribution (negative versus positive contribution on the outcome). For other contributors, such as group fairness and their associated KPIs, similar analysis and calculations have been performed with the aim of adding profundity to the proposed audit framework. The framework is expected to assist regulatory bodies in performing conformity assessments of AI systems using multilevel binomial classification models at businesses. It will also benefit providers, users, and assessment bodies, as defined in the European Commission’s proposed Regulation on Artificial Intelligence, when deploying AI-systems such as MLMs, to be future-proof and aligned with the regulation.

Exploiting Textual Information for Fake News Detection.

"Fake news" refers to the deliberate dissemination of news with the purpose to deceive and mislead the public. This paper assesses the accuracy of several Machine Learning (ML) algorithms, using a style-based technique that relies on textual information extracted from news, such as part of speech counts. To expand the already proposed styled-based techniques, a new method of enhancing a linguistic feature set is proposed. It combines Named Entity Recognition (NER) with the Frequent Pattern (FP) Growth association rule mining algorithm, aiming to provide better insight into the papers' sentence level structure. Recursive feature elimination was used to identify a subset of the highest performing linguistic characteristics, which turned out to align with the literature. Using pre-trained word embeddings, document embeddings and weighted document embeddings were constructed using each word's TF-IDF value as the weight factor. The document embeddings were mixed with the linguistic features providing a variety of training/test feature sets. For each model, the best performing feature set was identified and fine-tuned regarding its hyper parameters to improve accuracy. ML algorithms' results were compared with two Neural Networks: Convolutional Neural Network (CNN) and Long-Short-Term Memory (LSTM). The results indicate that CNN outperformed all other methods in terms of accuracy, when companied with pre-trained word embeddings, yet SVM performs almost the same with a wider variety of input feature sets. Although style-based technique scores lower accuracy, it provides explainable results about the author's writing style decisions. Our work points out how new technologies and combinations of existing techniques can enhance the style-based approach capturing more information.

Interpretable Machine Learning Model for Predicting Pathologic Complete Response in Patients with Rectal Adenocarcinoma Treated with Chemoradiation Therapy

<h3>Purpose/Objective(s)</h3> Following neoadjuvant chemoradiotherapy (nCRT), pathologic complete response (pCR) strongly influences the decision to proceed with surgery versus "watchful waiting" in rectal cancer (RC) patients. The purpose of this study is to predict pCR without using invasive procedures. An interpretable machine learning model trained with clinical and imaging data from diagnosis and treatment, with extracted radiomics (R) and dosiomics (D) features is utilized to gain insight into contributing factors. <h3>Materials/Methods</h3> This study used multi-institutional datasets, including a training set of 180 patients from our institution and an independent test set of 37 patients from the RTOG 0822 clinical trial. Each patient had a radiotherapy planning CT and the associated contours of the gross tumor volume (GTV) and the organ-at-risks (OARs) including the bladder, bowel_samll, and femur_heads. A total of 296 features including clinical parameters (CP), GTV and OAR dose-volume histogram (DVH), GTV R, and GTV D features were extracted. R and D features were subcategorized into the first- (L1), second- (L2), and higher-order (L3) local texture features. Multiview input data analysis was performed to identify an optimal set of input feature categories by using an exhaustive search. For each input, feature selection was performed using Boruta, followed by collinearity removal based on the variance inflation factor. Explainable boosting machine (EBM), an interpretable glass-box model, was trained using selected features. The performance of EBM on the test set was evaluated using the area under the receiver operating characteristic curve (AUC) and compared with that of 3 state-of-the-art black-box models: extreme gradient boosting (XGB), random forest (RF), and support vector machine (SVM). <h3>Results</h3> Selected features included two shape (elongation, maximum2DDiameterColumn) and one first-order (variance) features in R. Global explanations of EBM showed that tumors with maximum2DDiameterColumn <40 or >90 mm, elongation <0.55, and higher variance of CT intensities were associated with unfavorable outcomes. For all 4 models, the best predictive performance was obtained with an optimal input feature subset of CP+DVH+L1_R+L1_D, which significantly outperformed a full feature set of CP+DVH+R_all+D_all (AUC=0.602 to 0.618). The EBM had the best performance for predicting pCR (AUC=0.855), followed by RF, SVM, and XGB (AUC=0.839, 0.822, and 0.809). <h3>Conclusion</h3> EBM has the potential to improve the predictability of pCR while also enhancing model interpretability, which can aid in radiotherapy decision-making. Using an optimal input feature set selected by multi-view input analysis can lead to improved performance compared to concatenating all available features, which may, in turn, have implications in patient selection for a "watchful-waiting" approach in RC management.

Interpretable stock price forecasting model using genetic algorithm-machine learning regressions and best feature subset selection

Recent stock market studies adopting machine learning and deep learning techniques have achieved remarkable performances with convenient accessibility. However, machine learning and deep learning models are notorious for their black-box structure. To build human-friendly interpretability in stock price prediction, many studies focus on the relationship between input features and the outcome by measuring the feature importance. However, the feature-importance-based interpretability methods have such drawbacks as relative feature importance, vague importance of correlated features, and impractical interpretability. Furthermore, they overlook two principal characteristics of time series stock price data: time-dependency and collective behavior of features. As a solution to catch the collective behavior of features over a whole data period, we propose the best feature subset selection. Additionally, for the solution to reflect the time-dependent characteristic of stock price data over a short data period, we propose piecewise best feature subset selection. The proposed algorithm uses two separate input feature sets: internal technical indicators and external market prices. This bilateral forecasting scheme goes through a two-stage feature selection process composed of feature set expansion, hybridized genetic algorithm-machine learning regressions to select important features, and importance score filtering to select optimal features. Finally, the best feature subset is selected for forecasting and interpretation. The proposed method achieves the best feature subset of parsimoniously fewer features for interpretability and improves average forecasting Root Mean Squared Error by 10.42% for the optimal feature set and 13.47% for the best feature subset of the internal technical indicators. For enhanced local interpretability in this study, we use Savitzky–Golay smoothing as part of piecewise optimal curve fitting to examine each potential grouping of external features. The proposed local interpretability technique using piecewise optimal curve fitting and piecewise best feature subset provides a more timely-flexible interpretation of stock price behavior using a few best features for piecewise data segments. Compared with other feature-importance interpretability techniques that only rely on either a single data point or a whole data period, the proposed interpretability technique overcomes their limitations, and reflects the main characteristics of time series data.

Modelling patient trajectories using multimodal information

Background:Electronic Health Records (EHRs) aggregate diverse information at the patient level, holding a trajectory representative of the evolution of the patient health status throughout time. Although this information provides context and can be leveraged by physicians to monitor patient health and make more accurate prognoses/diagnoses, patient records can contain information from very long time spans, which combined with the rapid generation rate of medical data makes clinical decision making more complex. Patient trajectory modelling can assist by exploring existing information in a scalable manner, and can contribute in augmenting health care quality by fostering preventive medicine practices (e.g. earlier disease diagnosis). Methods:We propose a solution to model patient trajectories that combines different types of information (e.g. clinical text, standard codes) and considers the temporal aspect of clinical data. This solution leverages two different architectures: one supporting flexible sets of input features, to convert patient admissions into dense representations; and a second exploring extracted admission representations in a recurrent-based architecture, where patient trajectories are processed in sub-sequences using a sliding window mechanism. Results:The developed solution was evaluated on two different clinical outcomes, unexpected patient readmission and disease progression, using the publicly available Medical Information Mart for Intensive Care (MIMIC)-III clinical database. The results obtained demonstrate the potential of the first architecture to model readmission and diagnoses prediction using single patient admissions. While information from clinical text did not show the discriminative power observed in other existing works, this may be explained by the need to fine-tune the clinicalBERT model. Finally, we demonstrate the potential of the sequence-based architecture using a sliding window mechanism to represent the input data, attaining comparable performances to other existing solutions. Conclusion:Herein, we explored DL-based techniques to model patient trajectories and propose two flexible architectures that explore patient admissions on an individual and sequence basis. The combination of clinical text with other types of information led to positive results, which can be further improved by including a fine-tuned version of clinicalBERT in the architectures. The proposed solution can be publicly accessed at https://github.com/bioinformatics-ua/PatientTM.