Build Machine Learning Models Research Articles

Predicting physiological responses of dairy cows using comprehensive variables

Heat stress is increasingly affecting the production, health, and reproduction of dairy cows. Previous studies used limited variables as predictors of physiological responses, and the developed models poorly predict animal responses in evaporatively cooled environments. The aim of this study was to build machine learning models using comprehensive variables to predict physiological responses of dairy cows raised on an actual dairy farm equipped with sprinklers. Four algorithms including random forests, gradient boosting machines, artificial neural networks (ANN), and regularized linear regression were used to predict respiration rate (RR), vaginal temperature (VT), and eye temperature (ET) with 13 predictor variables from three dimensions: production, cow-related, and environmental factors. The classification performance of the predicted values in recognizing individual heat stress states was compared with commonly used thermal indices. The performance on the testing sets shows that the ANN models yielded the lowest root mean squared error for predicting RR (13.24 breaths/min), VT (0.30 °C), and ET (0.29 °C). The results interpreted with partial dependence plots and Local Interpretable Model-agnostic Explanations show that P.M. measurements and winter calving contributed most to high RR and VT predictions, whereas lying posture, high ambient temperature, and low wind speed contributed most to high ET predictions. When determining the ground-truth heat stress state by the actual RR, the best classification performance was yielded by the predicted RR with an accuracy of 77.7%; when determining the ground-truth heat stress state by the actual VT, the best classification performance was yielded by the predicted VT with an accuracy of 75.3%. This study demonstrates the ability of ANN in predicting physiological responses of dairy cows raised on actual farms with access to sprinklers. Adding more predictors other than meteorological parameters into training could increase predictive performance. Recognizing the heat stress state of individual animals, especially those at the highest risk, based on the predicted physiological responses and their interpretations can inform better heat abatement decisions.

Predicting drug shortages using pharmacy data and machine learning

Drug shortages are a global and complex issue having negative impacts on patients, pharmacists, and the broader health care system. Using sales data from 22 Canadian pharmacies and historical drug shortage data, we built machine learning models predicting shortages for the majority of the drugs in the most-dispensed interchangeable groups in Canada. When breaking drug shortages into four classes (none, low, medium, high), we were able to correctly predict the shortage class with 69% accuracy and a kappa value of 0.44, one month in advance, without access to any inventory data from drug manufacturers and suppliers. We also predicted 59% of the shortages deemed to be most impactful (given the demand for the drugs and the potential lack of interchangeable options). The models consider many variables, including the average days of a drug supply per patient, the total days of a drug supply, previous shortages, and the hierarchy of drugs within different drug groups and therapeutic classes. Once in production, the models will allow pharmacists to optimize their orders and inventories, and ultimately reduce the impact of drug shortages on their patients and operations.

Sex-Based Performance Disparities in Machine Learning Algorithms for Cardiac Disease Prediction: Exploratory Study.

The presence of bias in artificial intelligence has garnered increased attention, with inequities in algorithmic performance being exposed across the fields of criminal justice, education, and welfare services. In health care, the inequitable performance of algorithms across demographic groups may widen health inequalities. Here, we identify and characterize bias in cardiology algorithms, looking specifically at algorithms used in the management of heart failure. Stage 1 involved a literature search of PubMed and Web of Science for key terms relating to cardiac machine learning (ML) algorithms. Papers that built ML models to predict cardiac disease were evaluated for their focus on demographic bias in model performance, and open-source data sets were retained for our investigation. Two open-source data sets were identified: (1) the University of California Irvine Heart Failure data set and (2) the University of California Irvine Coronary Artery Disease data set. We reproduced existing algorithms that have been reported for these data sets, tested them for sex biases in algorithm performance, and assessed a range of remediation techniques for their efficacy in reducing inequities. Particular attention was paid to the false negative rate (FNR), due to the clinical significance of underdiagnosis and missed opportunities for treatment. In stage 1, our literature search returned 127 papers, with 60 meeting the criteria for a full review and only 3 papers highlighting sex differences in algorithm performance. In the papers that reported sex, there was a consistent underrepresentation of female patients in the data sets. No papers investigated racial or ethnic differences. In stage 2, we reproduced algorithms reported in the literature, achieving mean accuracies of 84.24% (SD 3.51%) for data set 1 and 85.72% (SD 1.75%) for data set 2 (random forest models). For data set 1, the FNR was significantly higher for female patients in 13 out of 16 experiments, meeting the threshold of statistical significance (-17.81% to -3.37%; P<.05). A smaller disparity in the false positive rate was significant for male patients in 13 out of 16 experiments (-0.48% to +9.77%; P<.05). We observed an overprediction of disease for male patients (higher false positive rate) and an underprediction of disease for female patients (higher FNR). Sex differences in feature importance suggest that feature selection needs to be demographically tailored. Our research exposes a significant gap in cardiac ML research, highlighting that the underperformance of algorithms for female patients has been overlooked in the published literature. Our study quantifies sex disparities in algorithmic performance and explores several sources of bias. We found an underrepresentation of female patients in the data sets used to train algorithms, identified sex biases in model error rates, and demonstrated that a series of remediation techniques were unable to address the inequities present.

Machine-learning with region-level radiomic and dosimetric features for predicting radiotherapy-induced rectal toxicities in prostate cancer patients

Background and purposeThis study aims to build machine learning models to predict radiation-induced rectal toxicities for three clinical endpoints and explore whether the inclusion of radiomic features calculated on radiotherapy planning computerised tomography (CT) scans combined with dosimetric features can enhance the prediction performance. Materials and methods183 patients recruited to the VoxTox study (UK-CRN-ID-13716) were included. Toxicity scores were prospectively collected after 2 years with grade ≥ 1 proctitis, haemorrhage (CTCAEv4.03); and gastrointestinal (GI) toxicity (RTOG) recorded as the endpoints of interest. The rectal wall on each slice was divided into 4 regions according to the centroid, and all slices were divided into 4 sections to calculate region-level radiomic and dosimetric features. The patients were split into a training set (75%, N = 137) and a test set (25%, N = 46). Highly correlated features were removed using four feature selection methods. Individual radiomic or dosimetric or combined (radiomic + dosimetric) features were subsequently classified using three machine learning classifiers to explore their association with these radiation-induced rectal toxicities. ResultsThe test set area under the curve (AUC) values were 0.549, 0.741 and 0.669 for proctitis, haemorrhage and GI toxicity prediction using radiomic combined with dosimetric features. The AUC value reached 0.747 for the ensembled radiomic-dosimetric model for haemorrhage. ConclusionsOur preliminary results show that region-level pre-treatment planning CT radiomic features have the potential to predict radiation-induced rectal toxicities for prostate cancer. Moreover, when combined with region-level dosimetric features and using ensemble learning, the model prediction performance slightly improved.

Using machine learning on tree‐ring data to determine the geographical provenance of historical construction timbers

AbstractDendroclimatology offers the unique opportunity to reconstruct past climate at annual resolution and wood from historical buildings can be used to extend such information back in time up to several millennia. However, the varying and often unclear origin of timbers affects the climate sensitivity of individual tree‐ring samples. Here, we compare tree‐ring width and density of 143 living larch (Larix decidua Mill.) trees at seven sites along an elevational transect from 1400 to 2200 m asl and 99 historical tree‐ring series to parametrize state‐of‐the‐art classification models for the European Alps. To achieve geographical provenance of the historical series, nine different supervised machine learning algorithms are trained and tested in their capability to solve our classification problem. Based on this assessment, we consider a tree‐ring density‐based and a tree‐ring width‐based dataset for model building. For each of these datasets, a general not species‐related model and a larch‐specific model including the cyclic larch budmoth influence are built. From the nine tested machine learning algorithms, Extreme Gradient Boosting showed the best performance. The density‐based models outperform the ring‐width models with the larch‐specific density model reaching the highest skill (f1 score = 0.8). The performance metrics reveal that the larch‐specific density model also performs best within individual sites and particularly in sites above 2000 m asl, which show the highest temperature sensitivities. The application of the specific density model for larch allows the historical series to be assigned with high confidence to a particular elevation within the valley. The procedure can be applied to other provenance studies using multiple tree growth characteristics. The novel approach of building machine learning models based on tree‐ring density features allows to omit a common period between reference and historical data for finding the provenance of relict wood and will therefore help to improve millennium‐length climate reconstructions.

SCA: Sybil-Based Collusion Attacks of IIoT Data Poisoning in Federated Learning

With the massive amounts of data generated by industrial Internet of Things (IIoT) devices at all moments, federated learning (FL) enables these distributed distrusted devices to collaborate to build machine learning model while maintaining data privacy. However, malicious participants still launch malicious attacks against the security vulnerabilities during model aggregation. This article is the first to propose Sybil-based collusion attacks (SCA) in the IIoT-FL system for the vulnerabilities mentioned above. The malicious participants use label flipping attacks to complete local poisoning training. Meanwhile, they can virtualize multiple Sybil nodes to make the local poisoning models aggregated with the greatest possibility during aggregation. They focus on making the joint model misclassify the selected attack class samples during the testing phase, while other nonattack classes kept the main task accuracy similar to the nonpoisoned state. Exhaustive experimental analysis demonstrates that our SCA has a superior performance on multiple aspects than the state-of-the-art.

Abstract P2-12-02: ClinicalomicsDB - Bridging the gap between clinical omics data and machine learning

Abstract BACKGROUND: Clinical trials are controlled patient studies aiming to objectively assess the effectiveness of treatment interventions. However, the average effectiveness observed at the group level does not directly apply to individual patients. Cancer clinical trials that include molecular profiling on the baseline tumor samples enable the discovery of treatment response-associated molecular features and the development of prediction models that help discriminate responders from non-responders in each treatment arm. Despite potential opportunities, there are many challenges associated with the predictive modeling of such data, as exemplified by the small sample size and large feature space. Advances in machine learning and artificial intelligence (AI) may provide new solutions to address these challenges. However, for machine learning and AI to have an impact, data needs to be carefully curated, high-quality, standardized, and easily accessible and understood by data scientists, who may not have the domain knowledge. METHODS: Here, we created a python package “ClinicalOmicsDB” to address the challenges of data accessibility and promote development and application of machine learning methods to omics data from clinical trial samples with treatment response information. The package makes data readily analyzable by data scientists so that they can develop, utilize, and optimize their algorithms for predicting treatment responses and discovering novel biomarkers. To promote a two-way dialogue, we have also developed several Jupyter Notebook tutorials for biologists or clinicians who wish to gain expertise in machine learning. Omics data from clinical studies are downloaded from Gene Expression Omnibus (GEO) and responses were determined based on clinical trial primary endpoints. Currently, the package has datasets from 22 breast cancer clinical trials, including a total of 5050 patients (Table 1). It will be continuously expanded to include additional trials for breast cancer and other cancer types. RESULTS: To evaluate package utility, we built machine learning models to predict neoadjuvant chemotherapy with four cycles of 5-fluorouracil/epirubicin/cyclophosphamide (FEC) followed by four cycles of docetaxel/capecitabine on US Oncology clinical trial 02-103 [GSE42822]. The best performing model was the Random Forest Classifier model, which had an AUC of 0.817. To determine the generalizability of machine learning models established from the package, we trained a Random Forest Classifier model using the GSE25055 breast cancer dataset and apply the model to a different breast cancer dataset, GSE20194, which yielded an AUC of 0.648. These results suggest utilizing machine learning on clinical omics datasets can provide predictive and generalizable models that could be implemented in clinical settings for future breast cancer patients. CONCLUSION: We are expanding the database for data scientists, biologists, and clinicians to practice novel biotechnology-derived therapies to facilitate the implementation of precision medicine approaches for future patients. As more people add new data to the package, we will work towards improving pharmaceutical and private companies’ clinical trial data sharing policies and practices to promote data sharing. Table 1. Available breast cancer datasets in ClinicalOmicsDB The current database has 22 breast cancer clinical trials with 5050 total patients. Therapy shows various cytotoxic and/or targeted treatments utilized in a clinical trial. The database will be continuously expanded to include additional trials for breast cancer and other cancer types. Citation Format: Chang In Moon, Byron Jia, Bing Zhang. ClinicalomicsDB - Bridging the gap between clinical omics data and machine learning [abstract]. In: Proceedings of the 2022 San Antonio Breast Cancer Symposium; 2022 Dec 6-10; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2023;83(5 Suppl):Abstract nr P2-12-02.

A Comprehensive and Improved Definition for Hospital-Acquired Pressure Injury Classification Based on Electronic Health Records: Comparative Study

Patients develop pressure injuries (PIs) in the hospital owing to low mobility, exposure to localized pressure, circulatory conditions, and other predisposing factors. Over 2.5 million Americans develop PIs annually. The Center for Medicare and Medicaid considers hospital-acquired PIs (HAPIs) as the most frequent preventable event, and they are the second most common claim in lawsuits. With the growing use of electronic health records (EHRs) in hospitals, an opportunity exists to build machine learning models to identify and predict HAPI rather than relying on occasional manual assessments by human experts. However, accurate computational models rely on high-quality HAPI data labels. Unfortunately, the different data sources within EHRs can provide conflicting information on HAPI occurrence in the same patient. Furthermore, the existing definitions of HAPI disagree with each other, even within the same patient population. The inconsistent criteria make it impossible to benchmark machine learning methods to predict HAPI. The objective of this project was threefold. We aimed to identify discrepancies in HAPI sources within EHRs, to develop a comprehensive definition for HAPI classification using data from all EHR sources, and to illustrate the importance of an improved HAPI definition. We assessed the congruence among HAPI occurrences documented in clinical notes, diagnosis codes, procedure codes, and chart events from the Medical Information Mart for Intensive Care III database. We analyzed the criteria used for the 3 existing HAPI definitions and their adherence to the regulatory guidelines. We proposed the Emory HAPI (EHAPI), which is an improved and more comprehensive HAPI definition. We then evaluated the importance of the labels in training a HAPI classification model using tree-based and sequential neural network classifiers. We illustrate the complexity of defining HAPI, with <13% of hospital stays having at least 3 PI indications documented across 4 data sources. Although chart events were the most common indicator, it was the only PI documentation for >49% of the stays. We demonstrate a lack of congruence across existing HAPI definitions and EHAPI, with only 219 stays having a consensus positive label. Our analysis highlights the importance of our improved HAPI definition, with classifiers trained using our labels outperforming others on a small manually labeled set from nurse annotators and a consensus set in which all definitions agreed on the label. Standardized HAPI definitions are important for accurately assessing HAPI nursing quality metric and determining HAPI incidence for preventive measures. We demonstrate the complexity of defining an occurrence of HAPI, given the conflicting and incomplete EHR data. Our EHAPI definition has favorable properties, making it a suitable candidate for HAPI classification tasks.

Federated Learning for Data Analytics in Education

Federated learning techniques aim to train and build machine learning models based on distributed datasets across multiple devices while avoiding data leakage. The main idea is to perform training on remote devices or isolated data centers without transferring data to centralized repositories, thus mitigating privacy risks. Data analytics in education, in particular learning analytics, is a promising scenario to apply this approach to address the legal and ethical issues related to processing sensitive data. Indeed, given the nature of the data to be studied (personal data, educational outcomes, and data concerning minors), it is essential to ensure that the conduct of these studies and the publication of the results provide the necessary guarantees to protect the privacy of the individuals involved and the protection of their data. In addition, the application of quantitative techniques based on the exploitation of data on the use of educational platforms, student performance, use of devices, etc., can account for educational problems such as the determination of user profiles, personalized learning trajectories, or early dropout indicators and alerts, among others. This paper presents the application of federated learning techniques to a well-known learning analytics problem: student dropout prediction. The experiments allow us to conclude that the proposed solutions achieve comparable results from the performance point of view with the centralized versions, avoiding the concentration of all the data in a single place for training the models.

Text embedding techniques for efficient clustering of twitter data.

World wide web is abundant with various types of information such blogs, social media posts, news articles. With this type of magnitude of online content, there is a need to deeply understand the insights of it in order to make use of the information for practical applications such as event detection, polarity, sentiment analysis and so on. Natural Language Processing (NLP) is the study of such information which is used for text classification, sentiment analysis, clustering of similar text. NLP makes use of linguistic knowledge and build machine learning models to analyse textual information. NLP finds its way in various applications like classification of online review into positive and negative without actually reading the reviews and feedback. For text analysis, there should be a way to quantify the text based on its frequency of occurrence, correlation with neighbouring words, contextual similarity of words, etc. One such way is word embedding. This study applies various word embedding techniques on tweets of popular news channels and clusters the resultant vectors using K-means algorithm. From this study, it is found out that Bidirectional Encoder Representations from Transformers (BERT) has achieved highest accuracy rate when used with K-means clustering.

Systematic Evaluation of Local and Global Machine Learning Models for the Prediction of ADME Properties.

Machine learning (ML) has become an indispensable tool to predict absorption, distribution, metabolism, and excretion (ADME) properties in pharmaceutical research. ML algorithms are trained on molecular structures and corresponding ADME assay data to develop quantitative structure-property relationship (QSPR) models. Traditional QSPR models were trained on compound sets of limited size. With the advent of more complex ML algorithms and data availability, training sets have become larger and more diverse. Most common training approaches consist in either training a model with a small set of similar compounds, namely, compounds designed for the same drug discovery project or chemical series (local model approach) or with a larger set of diverse compounds (global model approach). Global models are built with all experimental data available for an assay, combining compound data from different projects and disease areas. Despite the ML progress made so far, the choice of the appropriate data composition for building ML models is still unclear. Herein, a systematic evaluation of local and global ML models was performed for 10 different experimental assays and 112 drug discovery projects. Results show a consistent superior performance of global models for ADME property predictions. Diagnostic analyses were also carried out to investigate the influence of training set size, structural diversity, and data shift in the relative performance of local and global ML models. Training set and structural diversity did not have an impact in the relative performance on the methods. Instead, data shift helped to identify the projects with larger performance differences between local and global models. Results presented in this work can be leveraged to improve ML-based ADME properties predictions and thus decision-making in drug discovery projects.

Machine Learning for Early Outcome Prediction in Septic Patients in the Emergency Department

Background: Sepsis is one of the major causes of in-hospital death, and is frequent in patients presenting to the emergency department (ED). Early identification of high-risk septic patients is critical. Machine learning (ML) techniques have been proposed for identification and prognostication of ED septic patients, but these models often lack pre-hospital data and lack validation against early sepsis identification scores (such as qSOFA) and scores for critically ill patients (SOFA, APACHE II). Methods We conducted an electronic health record (EHR) study to test whether interpretable and scalable ML models predict mortality in septic ED patients and compared their performance with clinical scores. Consecutive adult septic patients admitted to ED over 18 months were included. We built ML models, ranging from a simple-classifier model, to unbalanced and balanced logistic regression, and random forest, and compared their performance to qSOFA, SOFA, and APACHE II scores. Results: We included 425 sepsis patients after screening 38,500 EHR for sepsis criteria. Overall mortality was 15.2% and peaked in patients coming from retirement homes (38%). Random forest, like balanced (0.811) and unbalanced logistic regression (0.863), identified patients at risk of mortality (0.813). All ML models outperformed qSOFA, APACHE II, and SOFA scores. Age, mean arterial pressure, and serum sodium were major mortality predictors. Conclusions: We confirmed that random forest models outperform previous models, including qSOFA, SOFA, and APACHE II, in identifying septic patients at higher mortality risk, while maintaining good interpretability. Machine learning models may gain further adoption in the future with increasing diffusion and granularity of EHR data, yielding the advantage of increased scalability compared to standard statistical techniques.

Descriptive and predictive analysis identify centenarians' characteristics from the Basque population.

Centenarians exhibit extreme longevity and have been postulated, by some researchers, as a model for healthy aging. The identification of the characteristics of centenarians might be useful to understand the process of human aging. In this retrospective study, we took advantage of demographic, clinical, biological, and functional data of deceased individuals between 2014 and 2020 in Guipúzcoa (Basque Country, Spain) taken from the Basque Health Service electronic health records data lake. Fifty characteristics derived from demographic, clinical, pharmaceutical, biological, and functional data were studied in the descriptive analysis and compared through differences in means tests. Twenty-seven of them were used to build machine learning models in the predictive analysis and their relevance for classifying centenarians was assessed. Most centenarians were women and lived in nursing homes. Importantly, they developed fewer diseases, took fewer drugs, and required fewer medical attendances. They also showed better biological profiles, exhibiting lower levels of glucose, hemoglobin, glycosylated hemoglobin, and triglycerides in blood analysis compared with non-centenarians. In addition, machine learning analyses revealed the main characteristics of the profiles associated with centenarians' status as being women, having fewer consultations, having fewer diagnoses of neoplasms, and having lower levels of hemoglobin. Our results revealed the main characteristics linked to centenarians in the Basque Country using Computational Biology programs. These results expand the knowledge on the characterization of the centenarian population and hence of human longevity.

How to use no-code artificial intelligence to predict and minimize the inventory distortions for resilient supply chains

Consumers’ dramatic demand has a pernicious effect throughout the supply chain. It exacerbates inventory distortion because of significant revenue loss caused by stock-level issues. Despite the availability of several forecasting techniques, large organisations, manufacturing firms, and e-commerce websites collectively lose around $1.8 trillion annually to inventory distortion. If this problem is solved, sales may increase by 10.3 percent. The businesses are concerned about mitigating this loss. Artificial intelligence (AI) can play a significant role in building resilient supply chains. However, developing AI models consumes time and cost. In this paper, we propose a No Code Artificial Intelligence (NCAI) enabling non-technical companies to build machine learning models based on production quantity and inventory replenishment. The development of the NCAI model is fast and inexpensive. However, little research deals with applying NCAI to operations and supply chain problems. Addressing the existing gap, we show the application of NCAI in the retail industry.

Rationing scarce healthcare capacity: A study of the ventilator allocation guidelines during the COVID-19 pandemic.

In the United States, even though national guidelines for allocating scarce healthcare resources are lacking, 26 states have specific ventilator allocation guidelines to be invoked in case of a shortage. While several states developed their guidelines in response to the recent COVID-19 pandemic, New York State developed these guidelines in 2015 as "pandemic influenza is a foreseeable threat, one that we cannot ignore." The primary objective of this study is to assess the existing procedures and priority rules in place for allocating/rationing scarce ventilator capacity and propose alternative (and improved) priority schemes. We first build machine learning models using inpatient records of COVID-19 patients admitted to New York-Presbyterian/Columbia University Irving Medical Center and an affiliated community health center to predict survival probabilities as well as ventilator length-of-use. Then, we use the resulting point estimators and their uncertainties as inputs for a multiclass priority queueing model with abandonments to assess three priority schemes: (i) SOFA-P (Sequential Organ Failure Assessment based prioritization), which most closely mimics the existing practice by prioritizing patients with sufficiently low SOFA scores; (ii) ISP (incremental survival probability), which assigns priority based on patient-level survival predictions; and (iii) ISP-LU (incremental survival probability per length-of-use), which takes into account survival predictions and resource use duration. Our findings highlight that our proposed priority scheme, ISP-LU, achieves a demonstrable improvement over the other two alternatives. Specifically, the expected number of survivals increases and death risk while waiting for ventilator use decreases. We also show that ISP-LU is a robust priority scheme whose implementation yields a Pareto-improvement over both SOFA-P and ISP in terms of maximizing saved lives after mechanical ventilation while limiting racial disparity in access to the priorityqueue.

Brittle and ductile characteristics of intermetallic compounds in magnesium alloys: A large-scale screening guided by machine learning

In the present work, we have employed machine learning (ML) techniques to evaluate ductile-brittle (DB) behaviors in intermetallic compounds (IMCs) which can form magnesium (Mg) alloys. This procedure was mainly conducted by a proxy-based method, where the ratio of shear (G)/bulk (B) moduli was used as a proxy to identify whether the compound is ductile or brittle. Starting from compounds information (composition and crystal structure) and their moduli, as found in open databases (AFLOW), ML-based models were built, and those models were used to predict the moduli in other compounds, and accordingly, to foresee the ductile-brittle behaviors of these new compounds. The results reached in the present work showed that the built models can effectively catch the elastic moduli of new compounds. This was confirmed through moduli calculations done by density functional theory (DFT) on some compounds, where the DFT calculations were consistent with the ML prediction. A further confirmation on the reliability of the built ML models was considered through relating between the DB behavior in MgBe 13 and MgPd 2 , as evaluated by the ML-predicted moduli, and the nature of chemical bonding in these two compounds, which in turn, was investigated by the charge density distribution (CDD) and electron localization function (ELF) obtained by DFT methodology. The ML-evaluated DB behaviors of the two compounds was also consistent with the DFT calculations of CDD and ELF. These findings and confirmations gave legitimacy to the built model to be employed in further prediction processes. Indeed, as examples, the DB characteristics were investigated in IMCs that might from in three Mg alloy series, involving AZ, ZX and WE.

Automated credit assessment framework using ETL process and machine learning.

In the current business scenario, real-time analysis of enterprise data through Business Intelligence (BI) is crucial for supporting operational activities and taking any strategic decision. The automated ETL (extraction, transformation, and load) process ensures data ingestion into the data warehouse in near real-time, and insights are generated through the BI process based on real-time data. In this paper, we have concentrated on automated credit risk assessment in the financial domain based on the machine learning approach. The machine learning-based classification techniques can furnish a self-regulating process to categorize data. Establishing an automated credit decision-making system helps the lending institution to manage the risks, increase operational efficiency and comply with regulators. In this paper, an empirical approach is taken for credit risk assessment using logistic regression and neural network classification method in compliance with Basel II standards. Here, Basel II standards are adopted to calculate the expected loss. The required data integration for building machine learning models is done through an automated ETL process. We have concluded this research work by evaluating this new methodology for credit risk assessment.

Confound-leakage: confound removal in machine learning leads to leakage.

Machine learning (ML) approaches are a crucial component of modern data analysis in many fields, including epidemiology and medicine. Nonlinear ML methods often achieve accurate predictions, for instance, in personalized medicine, as they are capable of modeling complex relationships between features and the target. Problematically, ML models and their predictions can be biased by confounding information present in the features. To remove this spurious signal, researchers often employ featurewise linear confound regression (CR). While this is considered a standard approach for dealing with confounding, possible pitfalls of using CR in ML pipelines are not fully understood. We provide new evidence that, contrary to general expectations, linear confound regression can increase the risk of confounding when combined with nonlinear ML approaches. Using a simple framework that uses the target as a confound, we show that information leaked via CR can increase null or moderate effects to near-perfect prediction. By shuffling the features, we provide evidence that this increase is indeed due to confound-leakage and not due to revealing of information. We then demonstrate the danger of confound-leakage in a real-world clinical application where the accuracy of predicting attention-deficit/hyperactivity disorder is overestimated using speech-derived features when using depression as a confound. Mishandling or even amplifying confounding effects when building ML models due to confound-leakage, as shown, can lead to untrustworthy, biased, and unfair predictions. Our expose of the confound-leakage pitfall and provided guidelines for dealing with it can help create more robust and trustworthy ML models.

Toehold-Mediated Strand Displacement in Random Sequence Pools.

Toehold-mediated strand displacement (TMSD) has been used extensively for molecular sensing and computing in DNA-based molecular circuits. As these circuits grow in complexity, sequence similarity between components can lead to cross-talk, causing leak, altered kinetics, or even circuit failure. For small non-biological circuits, such unwanted interactions can be designed against. In environments containing a huge number of sequences, taking all possible interactions into account becomes infeasible. Therefore, a general understanding of the impact of sequence backgrounds on TMSD reactions is of great interest. Here, we investigate the impact of random DNA sequences on TMSD circuits. We begin by studying individual interfering strands and use the obtained data to build machine learning models that estimate kinetics. We then investigate the influence of pools of random strands and find that the kinetics are determined by only a small subpopulation of strongly interacting strands. Consequently, their behavior can be mimicked by a small collection of such strands. The equilibration of the circuit with the background sequences strongly influences this behavior, leading to up to 1 order of magnitude difference in reaction speed. Finally, we compare two established and one novel technique that speed up TMSD reactions in random sequence pools: a three-letter alphabet, protection of toeholds by intramolecular secondary structure, or by an additional blocking strand. While all of these techniques were useful, only the latter can be used without sequence constraints. We expect that our insights will be useful for the construction of TMSD circuits that are robust to molecular noise.

A prospective observational study for a Federated Artificial Intelligence solution for moniToring mental Health status after cancer treatment (FAITH): study protocol

BackgroundDepression is a common condition among cancer patients, across several points in the disease trajectory. Although presenting higher prevalence rates than the general population, it is often not reported or remains unnoticed. Moreover, somatic symptoms of depression are common in the oncological context and should not be dismissed as a general symptom of cancer. It becomes even more challenging to track psychological distress in the period after the treatment, where connection with the healthcare system typically becomes sporadic. The main goal of the FAITH project is to remotely identify and predict depressive symptoms in cancer survivors, based on a federated machine learning (ML) approach, towards optimization of privacy.MethodsFAITH will remotely analyse depression markers, predicting their negative trends. These markers will be treated in distinct categories, namely nutrition, sleep, activity and voice, assessed in part through wearable technologies. The study will include 300 patients who have had a previous diagnosis of breast or lung cancer and will be recruited 1 to 5 years after the end of primary cancer. The study will be organized as a 12-month longitudinal prospective observational cohort study, with monthly assessments to evaluate depression symptoms and quality of life among cancer survivors. The primary endpoint is the severity of depressive symptoms as measured by the Hamilton Depression Rating Scale (Ham-D) at months 3, 6, 9 and 12. Secondary outcomes include self-reported anxiety and depression symptoms (HADS scale), and perceived quality of life (EORTC questionnaires), at baseline and monthly. Based on the predictive models gathered during the study, FAITH will also aim at further developing a conceptual federated learning framework, enabling to build machine learning models for the prediction and monitoring of depression without direct access to user’s personal data.DiscussionImprovements in the objectivity of psychiatric assessment are necessary. Wearable technologies can provide potential indicators of depression and anxiety and be used for biofeedback. If the FAITH application is effective, it will provide healthcare systems with a novel and innovative method to screen depressive symptoms in oncological settings.Trial registrationTrial ID: ISRCTN10423782. Date registered: 21/03/2022.