Tree-based Classification Models Research Articles

TOWARDS CLICK-THROUGH RATE PREDICTION IN ONLINE ADVERTISING

The click-through rate is one of the main indicators based of which decisions are made in online advertising. A generalized approach to forecasting the click-through rate is proposed, taking into account the specifics of the industry data and using appropriate methods of processing categorical data and tree-based classification models. The simulation results can be used in the subsequent stages of analysis of advertising campaigns.

Evidence for identification of acute myeloid leukemia using peripheral blood's infrared spectrum and logistic regression

Invasive damage, poor prognosis and survival are problems for the traditional diagnostic and therapeutic methods of acute myeloid leukemia (AML), especially for the middle-aged and elderly. High sensitivity, less pain and fast speed are demands of AML diagnosis or long-term regular monitoring. Infrared (IR) spectroscopy has the potential to compensate for existing deficiencies and is explored in our work. Spectral differences between AML patients and controls are observed on the IR spectra, which indicate the changes in lipids and proteins in patients. The peak shift at 1390 cm−1 occurs on both peripheral blood (PB) and bone marrow (BM) spectra of AML patients making it possible to use PB rather than BM as the diagnostic source, which can relieve the suffering of diagnosis and monitoring. The logistic regression-based classification model can perfectly identify AML patients and controls with an accuracy of 100 % and the area under curve is equal to 1.00. Besides, some markers for identifying AML patients and controls are found. Statistical analysis shows that the peak area ratio A1451/A1390 has a significant difference, and the decision tree-based classification model finds that on the first derivative spectrum, the position of the largest difference is at 1391 cm−1. Furthermore, the proposed criterion based on markers we found can distinguish AML patients and controls with high accuracy of 95 %. Results show that PB’s IR spectrum combined with classification models can monitor the molecular-level biochemical changes and accurately identify AML patients and controls, which is helpful for the effective and efficient diagnosis and prognosis of AML.

Differentiating patients admitted primarily due to coronavirus disease 2019 (COVID-19) from those admitted with incidentally detected severe acute respiratory syndrome corona-virus type 2 (SARS-CoV-2) at hospital admission: A cohort analysis of German hospital records.

The number of hospitalized patients with severe acute respiratory syndrome coronavirus type 2 (SARS-CoV-2) does not differentiate between patients admitted due to coronavirus disease 2019 (COVID-19) (ie, primary cases) and incidental SARS-CoV-2 infection (ie, incidental cases). We developed an adaptable method to distinguish primary cases from incidental cases upon hospital admission. Retrospective cohort study. Data were obtained from 3 German tertiary-care hospitals. The study included patients of all ages who tested positive for SARS-CoV-2 by a standard quantitative reverse-transcription polymerase chain reaction (RT-PCR) assay upon admission between January and June 2022. We present 2 distinct models: (1) a point-of-care model that can be used shortly after admission based on a limited range of parameters and (2) a more extended point-of-care model based on parameters that are available within the first 24-48 hours after admission. We used regression and tree-based classification models with internal and external validation. In total, 1,150 patients were included (mean age, 49.5±28.5 years; 46% female; 40% primary cases). Both point-of-care models showed good discrimination with area under the curve (AUC) values of 0.80 and 0.87, respectively. As main predictors, we used admission diagnosis codes (ICD-10-GM), ward of admission, and for the extended model, we included viral load, need for oxygen, leucocyte count, and C-reactive protein. We propose 2 predictive algorithms based on routine clinical data that differentiate primary COVID-19 from incidental SARS-CoV-2 infection. These algorithms can provide a precise surveillance tool that can contribute to pandemic preparedness. They can easily be modified to be used in future pandemic, epidemic, and endemic situations all over the world.

Empirical analysis of tree-based classification models for customer churn prediction

Customer churn is a vital and reoccurring problem facing most business industries, particularly the telecommunications industry. Considering the fierce competition among telecommunications firms and the high expenses of attracting and gaining new subscribers, keeping existing loyal subscribers becomes crucial. Early prediction of disgruntled subscribers can assist telecommunications firms in identifying the reasons for churn and in deploying applicable innovative policies to boost productivity, maintain market competitiveness, and reduce monetary damages. Controlling customer churn through the development of efficient and dependable customer churn prediction (CCP) solutions is imperative to attaining this goal. According to the outcomes of current CCP research, several strategies, including rule-based and machine-learning (ML) processes, have been proposed to handle the CCP phenomenon. However, the lack of flexibility and robustness of rule based CCP solutions is a fundamental shortcoming, and the lopsided distribution of churn datasets is deleterious to the efficacy of most traditional ML techniques in CCP. Regardless, ML-based CCP solutions have been reported to be more effective than other forms of CCP solutions. Unlike linear-based, instance-based, and function-based ML classifiers, tree-based ML classifiers are known to generate predictive models with high accuracy, high stability, and ease of interpretation. However, the deployment of tree-based classifiers for CCP is limited in most cases to the decision tree (DT) and random forest (RF). Hence, this research investigated the effectiveness of tree-based classifiers with diverse computational properties in CCP. Specifically, the CCP performances of diverse tree-based classifiers such as the single, ensemble, enhanced, and hybrid tree-based classifiers are investigated. Also, the effects of data quality problems such as the class imbalance problem (CIP) on the predictive performances of tree-based classifiers and their homogeneous ensemble variants on CCP were assessed. From the experimental results, it was observed that the investigated tree-based classifiers outperformed other forms of classifiers such as linear-based (Support Vector Machine (SVM)), instance-based (K-Nearest Neighbour (KNN)), Bayesian-based (Naïve Bayes (NB)) and function-based (MultiLayer Perceptron (MLP)) classifiers in most cases with or without the CIP. Also, it was observed that the CIP has a significant effect on the CCP performances of investigated tree-based classifiers, but the combination of a data sampling technique and a homogeneous ensemble method can be an effective solution to CIP and also generate efficient CCP models.

The consumer intention to use e-commerce applications in the post-pandemic era: a predictive approach study using a CHAID tree-based algorithm

PurposeThis study proposes a hierarchic segmentation that develops a tree-based classification model and classifies the cases into groups. This allows for the definition of e-commerce user profiles for each of the groups. Additionally, it facilitates the development of actions to improve the adoption of the online channel that is in such high demand in the current pandemic COVID-19 context.Design/methodology/approachRegarding the created segments, two extreme segments stand out due to their marked differences and high volume. Segment 3 with 23% of the sample is the group with the most predisposition to use the online channel and is characterised by a high level of trust, more habitual use in comparison with other groups and the belief that its use implies high performance, which indicates they believe it to be useful, quick and helpful for more an effective shopping experience. The other extreme is found in segment 7. This group makes up 17.7% of the total and is the most reluctant to use the online channel. These users are characterised by the complete opposite: they have a low level of trust in this channel. However, the effort expectancy is low, i.e. they consider that the adoption of the online channel does not involve many difficulties in its learning and use. Nevertheless, they use it less regularly than the others.FindingsBased on the conclusions reached in this study, in the current pandemic context in which consumer demand for online shopping channels for all types of products is on the rise, it is recommended that companies focus on the following aspects. It is essential to build trust with the user and show them the real benefits of e-commerce, how it would improve their life and why they should use it. Additionally, it is vital that the user perceives it as an easy procedure that does not require a significant learning curve. Other fundamental aspects would be to reduce any uncertainty the user might have about the online shopping process, to make it as easy as possible, and to design a simple, intuitive and user-friendly interface. It is also recommendable to manage data usage efficiently. To do so, the authors recommend asking the user for the least amount of information possible, offering a data protection policy and assuring them that their information will not be misused nor shared with third parties. All of this provides a series of facilities to modify the online shopping habits of users. Research limitations/implicationsAs in most of the research, this study presents a series of limitations that should be debated and that could open future lines of investigation. Firstly, regarding the sample used that was limited to two neighbouring countries with similar profiles a priori; it would be necessary to compare their possible cultural differences according to Hofstede's dimensions as well as increase the number of European countries being analysed to reach a more generalised conclusions. Secondly, the variables used are a combination of those derived from the UTAUT2 model and others suggested in the literature as decisive in technology adoption by users, in this sense other theories and variables could be incorporated to complete a more holistic model.Practical implicationsThis work contributes in a general way to (1) analysing the intention to use e-commerce platforms from a set of antecedents previously defined by their importance, after a period of economic and social restrictions derived from the pandemic; (2) determination of customer segments from the classification made by the CHAID analysis; (3) characterisation of the previously defined segments through the successive divisions that were proposed in the analysis carried out.Social implicationsOther fundamental aspects would be to reduce any uncertainty the user might have about the online shopping process to make it as easy as possible, and to design a simple, intuitive, and user-friendly interface. It is also recommended to manage data usage efficiently. To do so, the authors recommend asking the user for the least amount of information possible, offering a data protection policy, and assuring them that their information will not be misused or shared with third parties.Originality/valueThe results obtained have allowed us to establish predictive and explanatory models of the behaviour of the segments and profiles created, which will help companies to improve their relationships with online customers in the coming years.

Impact of land tenure on livelihoods of women livestock-keepers in Burkina Faso

Land is an essential asset for women livestock-keepers that help to increase their livelihoods. The main objective of this study was to assess the impact of land tenure on production systems and the income of women livestock-keeper in Burkina Faso. A face to face survey was conducted on women’s livestock-keeping practices, their access to land and production using semi structured questionnaires. Hierarchical Cluster Analysis was used to assess the women’s livestock production systems. Conditional inference tree–based classification models allowed us to identify factors that mostly determine women’s access to livestock grazing land. A generalized linear model was used to assess the effects of land access parameters on the monthly income. A total of five production systems were identified comprising of two young, two adult and one older women’s systems. Furthermore, the determinants of women’s access to land varied across regions. Overall, the monthly income, animal species kept, herd size, land use rights (e.g., land rental and monthly payment, free use of the land, inheritance, purchase) determined women access to land in Burkina Faso. Only older women widow (7.78%) own land inherited from their husbands while married women had mainly free use of the land previously used by their husbands or their own acquired land. The assessment of the impact of land parameters on monthly income highlighted that in Central-Western Region, the land size and whether women were the landowners or not, had a very high significant and positive influence on the monthly income (p = 0.000444). An increase of one square meter (m2) of land size increased income by USD $3.45 per month. The same trend was observed in Eastern Region where an increase of one m2 increased income by USD $ 8.38 per month. In South-Western region, parameters such as the land size, formal education, training and better knowledge of women's right to land access led to increased income. Based on the findings of this research, it is clear that education as well as access to land are important factors that can improve women’s income from livestock production.

Machine learning-based soft-sensor development for road quality classification

Vibrations in road vehicles cause several harmful effects, health problems can occur for the passengers, and mechanical damage can occur to the vehicle components. Given the health, safety, and financial issues that arise, keeping the road network in good condition and detecting road defects as early as possible requires an extensive monitoring system. Related to this, our study presents the development of hardware and software for a low-cost, multi-sensor road quality monitoring system for passenger vehicles. The developed monitoring system can classify road sections according to their quality parameters into four classes. In order to detect vibrations in the vehicle, accelerometers and gyroscope sensors are installed at several points. Then, a machine learning-based soft-sensor development is introduced. Besides noise filtering, each data point is resampled by spatial frequency to reduce the velocity dependence. Subsequently, a decision tree-based classification model is trained using features from the power spectrum and principal component analysis. The classification algorithm is validated and tested with measurement data in a real-world environment. In addition to reviewing the accuracy of the model, we examine the correlation of the data measured in the cabin and on the suspension to see how much additional information is provided by the sensor on the axle.

Early prediction of severe retinopathy of prematurity requiring laser treatment using physiological data

BackgroundEarly risk stratification for developing retinopathy of prematurity (ROP) is essential for tailoring screening strategies and preventing abnormal retinal development. This study aims to examine the ability of physiological data during the first postnatal month to distinguish preterm infants with and without ROP requiring laser treatment.MethodsIn this cohort study, preterm infants with a gestational age <32 weeks and/or birth weight <1500 g, who were screened for ROP were included. Differences in the physiological data between the laser and non-laser group were identified, and tree-based classification models were trained and independently tested to predict ROP requiring laser treatment.ResultsIn total, 208 preterm infants were included in the analysis of whom 30 infants (14%) required laser treatment. Significant differences were identified in the level of hypoxia and hyperoxia, oxygen requirement, and skewness of heart rate. The best model had a balanced accuracy of 0.81 (0.72–0.87), a sensitivity of 0.73 (0.64–0.81), and a specificity of 0.88 (0.80–0.93) and included the SpO2/FiO2 ratio and baseline demographics (including gestational age and birth weight).ConclusionsRoutinely monitored physiological data from preterm infants in the first postnatal month are already predictive of later development of ROP requiring laser treatment, although validation is required in larger cohorts.ImpactRoutinely monitored physiological data from the first postnatal month are predictive of later development of ROP requiring laser treatment, although model performance was not significantly better than baseline characteristics (gestational age, birth weight, sex, multiple birth, prenatal glucocorticosteroids, route of delivery, and Apgar scores) alone.A balanced accuracy of 0.81 (0.72–0.87), a sensitivity of 0.73 (0.64–0.81), and a specificity of 0.88 (0.80–0.93) was achieved with a model including the SpO2/FiO2 ratio and baseline characteristics.Physiological data have potential to play a significant role for future ROP prediction and provide opportunities for early interventions to protect infants from abnormal retinal development.

Multifarious analytical capabilities of the UV/Vis protein fluorescence in blood plasma

Autofluorescence of blood plasma has been broadly considered as a prospective disease screening method. However, the assessment of such intrinsic fluorescence is mostly phenomenological, and its origin is still not fully understood, complicating its use in the clinical practice. Here we present the detailed evaluation of analytical capabilities, variability, and formation of blood plasma protein fluorescence based on the open dataset of excitation-emission matrices measured for ∼300 patients with suspected colorectal cancer, and our supporting model experiments. Using high-resolution size-exclusion chromatography coupled with comprehensive spectral analysis, we demonstrate, for the first time, the dominant role of HSA in the formation of blood plasma fluorescence in the visible spectral range (excitation wavelength >350 nm), presumably caused by its oxidative modifications. Furthermore, the diagnostic value of the tryptophan emission, as well as of the tyrosine fluorescence and visible fluorescence of proteins is shown by building a tree-based classification model that uses a small subset of physically interpretable fluorescence features for distinguishing between the control group and cancer patients with >80% accuracy. The obtained results extend current understanding and approaches used for the analysis of blood plasma fluorescence and pave the way for novel autofluorescence-based disease screening methods.

Unemployment in Rural Europe: A Machine Learning Perspective

This paper aims to provide policy-relevant findings that can contribute to the resilience of rural regions by discovering the main individual-level factors related to unemployment in those areas through the use of a set of machine learning techniques. Unemployment status is predicted using tree-based classification models: namely, classification tree, bootstrap aggregation, random forest, gradient boosting, and stochastic gradient boosting. The results are further analyzed using inferential techniques such as SHAP value analysis. Results suggest that access to training programmes can mitigate the labor market inequalities caused by differences in education levels, gender, age, alongside with parental education levels. The results also show how such inequalities are even larger for various subgroups detected by the employed algorithms.

Metagenomic analysis reveals gut bacterial signatures for diagnosis and treatment outcome prediction in bipolar depression

BackgroundWe aimed to characterize gut microbial alterations in depressed patients with bipolar disorder (BD) following quetiapine monotherapy and explored its potential for disease diagnosis and outcome prediction. MethodsFecal samples were obtained from 60 healthy individuals and 62 patients in acute depressive episodes. All patients received one-month quetiapine treatment after enrollment. The structure of gut microbiota was measured with metagenomic sequencing, and its correlation with clinical profiles and brain function as indicated by resting-state functional magnetic resonance imaging was analyzed. Random forest models based on bacterial species were constructed to distinguish patients from controls, and responders from non-responders, respectively. ResultsBD patients displayed specific alterations in gut microbial diversity and composition. Quetiapine treatment increased the diversity of microbial communities and changed the composition. The abundance of Clostridium bartlettii was negatively associated with age, baseline depression severity, while positively associated with spontaneous neural oscillation in the hippocampus. Tree-based classification models for (1) patients and controls and (2) responders and non-responders showed an area under the curve of 0.733 and 0.800, respectively. ConclusionOur findings add new evidence to the existing literature regarding gut dysbiosis in BD and reveal the potential of microbe-based biomarkers for disease diagnosis and treatment outcome prediction.

Mutational profiling of myeloid neoplasms associated genes may aid the diagnosis of acute myeloid leukemia with myelodysplasia-related changes

AML with myelodysplasia-related changes (AML-MRC) is a subtype of AML known to have adverse prognosis. The karyotype abnormalities in AML-MRC have been well established; however, relatively little has been known about the role of gene mutation profiles by next generation sequencing. 177 AML patients (72 AML-MRC and 105 non-MRC AML) were analyzed by NGS panel covering 53 AML related genes. AML-MRC showed statistically significantly higher frequency of TP53 mutation, but lower frequencies of mutations in NPM1, FLT3-ITDLow, FLT3-ITDHigh, FLT3-TKD, NRAS, and PTPN11 than non-MRC AML. Supervised tree-based classification models including Decision tree, Random forest, and XGboost, and logistic regression were used to evaluate if the mutation profiles could be used to aid the diagnosis of AML-MRC. All methods showed good accuracy in differentiating AML-MRC from non-MRC AML with AUC (area under curve) of ROC ranging from 0.69 to 0.78. Additionally, logistic regression indicated 3 independent factors (age and mutations of TP53 and FLT3) could aid the diagnosis AML-MRC. Using weighted factors, a AML-MRC risk scoring equation was established for potential application in clinical setting: +1x(Age ≥ 65) + 3 x (TP53 mutation) – 2 x (FLT3 mutation). Using a cutoff score of 0, the accuracy of the risk score was 0.76 with sensitivity of 0.77 and specificity of 0.75 for predicting the diagnosis of AML-MRC. Further studies with larger sample sizes are warranted to further evaluate the potential of using gene mutation profiles to aid the diagnosis of AML-MRC.

Effect of serum isoniazid level on treatment outcomes among tuberculosis patients with slow response – A retrospective cohort study

BackgroundIn this study, we investigate the effects of low serum TB drug level on treatment outcome among TB patients with slow response in South Korea, where the prevalence of rapid acetylator is relatively high. MethodsAmong the pulmonary TB patients whose treatment outcomes were reported between 2014 and 2018 at Incheon St. Mary hospital, those who underwent TDM because of delayed culture conversion or reversion were included. Primary outcome was microbiological failure defined as (1) positive sputum culture after 120 days of treatment, or (2) culture-confirmed relapse within one year after treatment completion. Patients with culture conversion within 120 days and no relapse were classified as the final conversion group. Clinical characteristics and serum drug concentration at 2 h after administration (C2hr) were compared between those two groups. ResultsA total of 55 pulmonary TB patients were included. Prevalence of subtherapeutic range of C2hr for isoniazid and rifampin was 78.2% and 21.8%, respectively. With one year of follow-up, 21 cases were classified as the microbiological failure group, and 34 cases as the final conversion group. In a multivariable logistic regression model for predicting microbiological failure, C2hr of isoniazid was the most significant predictor after adjusting for the effects of age and sex (adjusted odds ratio, 0.29; p = 0.009). In a tree-based classification model, C2hr of isoniazid with cutoff level 2.5 μg/ml was the most important variable for predicting microbiological failure. ConclusionsLow serum isoniazid level was related to poor treatment outcomes among the TB patients with slow response.

Forecasting the Walking Assistance Rehabilitation Level of Stroke Patients Using Artificial Intelligence.

Cerebrovascular accidents (CVA) cause a range of impairments in coordination, such as a spectrum of walking impairments ranging from mild gait imbalance to complete loss of mobility. Patients with CVA need personalized approaches tailored to their degree of walking impairment for effective rehabilitation. This paper aims to evaluate the validity of using various machine learning (ML) and deep learning (DL) classification models (support vector machine, Decision Tree, Perceptron, Light Gradient Boosting Machine, AutoGluon, SuperTML, and TabNet) for automated classification of walking assistant devices for CVA patients. We reviewed a total of 383 CVA patients’ (1623 observations) prescription data for eight different walking assistant devices from five hospitals. Among the classification models, the advanced tree-based classification models (LightGBM and tree models in AutoGluon) achieved classification results of over accuracy, recall, precision, and F1-score. In particular, AutoGluon not only presented the highest predictive performance (almost in accuracy, recall, precision, and F1-score, and in balanced accuracy) but also demonstrated that the classification performances of the tree-based models were higher than that of the other models on its leaderboard. Therefore, we believe that tree-based classification models have potential as practical diagnosis tools for medical rehabilitation.

Research shows scale of financial compensation, company culture and senior management performance have big influence on decisions to leave companies

PurposeThe purpose of this study is to develop tree-based binary classification models to predict the likelihood of employee attrition based on firm cultural and management attributes.Design/methodology/approachAfter preliminary analysis, the authors tested three hypotheses. One was: “Conditioned on an increase in salary, the magnitude of increase enhances the likelihood that an employee will leave their current firm despite differences between the old and new firm culture.” Two was: “Employees whose original firm has an overall rating greater than the 75th percentile that was also founded before 1900 are more likely to stay…” Finally, three was that: “Employees that maintain a low overall original firm rating are more likely to leave their firm upon a job transition, whereas those with higher overall ratings have a greater chance of remaining.”FindingsAfter analyzing thousands of online resumes submitted to Glassdoor’s portal, the authors found that the scale of financial compensation, the company culture and senior management performance all played a major role in influencing decisions to move on.Originality/valueThey offered three concrete recommendations based on the study. First, they said it was vital for companies to maintain strong Glassdoor.com ratings. The results revealed that firms in the top 10% of ratings were over 30% more likely to retain employees during a job transition than companies in the lowest 10%. Second, providing competitive salaries was necessary. Finally, the data showed a large discrepancy between senior management and CEO Glassdoor ratings. The researchers advised HR departments to closely monitor the impact of senior management behaviour.

Cost-Sensitive Ensemble Feature Ranking and Automatic Threshold Selection for Chronic Kidney Disease Diagnosis

Automated medical diagnosis is one of the important machine learning applications in the domain of healthcare. In this regard, most of the approaches primarily focus on optimizing the accuracy of classification models. In this research, we argue that, unlike general-purpose classification problems, medical applications, such as chronic kidney disease (CKD) diagnosis, require special treatment. In the case of CKD, apart from model performance, other factors such as the cost of data acquisition may also be taken into account to enhance the applicability of the automated diagnosis system. In this research, we proposed two techniques for cost-sensitive feature ranking. An ensemble of decision tree models is employed in both the techniques for computing the worth of a feature in the CKD dataset. An automatic threshold selection heuristic is also introduced which is based on the intersection of features’ worth and their accumulated cost. A set of experiments are conducted to evaluate the efficacy of the proposed techniques on both tree-based and non tree-based classification models. The proposed approaches were also evaluated against several comparative techniques. Furthermore, it is demonstrated that the proposed techniques select around 1/4th of the original CKD features while reducing the cost by a factor of 7.42 of the original feature set. Based on the extensive experimentation, it is concluded that the proposed techniques employing feature-cost interaction heuristic tend to select feature subsets that are both useful and cost-effective.

Achalasia subtypes can be identified with functional luminal imaging probe (FLIP) panometry using a supervised machine learning process.

Achalasia subtypes on high-resolution manometry (HRM) prognosticate treatment response and help direct management plan. We aimed to utilize parameters of distension-induced contractility and pressurization on functional luminal imaging probe (FLIP) panometry and machine learning to predict HRM achalasia subtypes. One hundred eighty adult patients with treatment-naïve achalasia defined by HRM per Chicago Classification (40 type I, 99 type II, 41 type III achalasia) who underwent FLIP panometry were included: 140 patients were used as the training cohort and 40 patients as the test cohort. FLIP panometry studies performed with 16-cm FLIP assemblies were retrospectively analyzed to assess distensive pressure and distension-induced esophageal contractility. Correlation analysis, single tree, and random forest were adopted to develop classification trees to identify achalasia subtypes. Intra-balloon pressure at 60mL fill volume, and proportions of patients with absent contractile response, repetitive retrograde contractile pattern, occluding contractions, sustained occluding contractions (SOC), contraction-associated pressure changes >10mmHg all differed between HRM achalasia subtypes and were used to build the decision tree-based classification model. The model identified spastic (type III) vs non-spastic (types I and II) achalasia with 90% and 78% accuracy in the train and test cohorts, respectively. Achalasia subtypes I, II, and III were identified with 71% and 55% accuracy in the train and test cohorts, respectively. Using a supervised machine learning process, a preliminary model was developed that distinguished type III achalasia from non-spastic achalasia with FLIP panometry. Further refinement of the measurements and more experience (data) may improve its ability for clinically relevant application.

Development of a screening tool to identify patients likely to benefit from clinical pharmacist review in a home-based primary care population

ObjectiveTo create a novel screening tool that identified patients who were most likely to benefit from pharmacist in-home medication reviews. DesignSingle-center, retrospective study. Setting and participantsA total of 25 homebound patients in Forsyth County, NC, aged 60 years or older with physical or cognitive impairments and enrolled in home-based primary care or transitional and supportive care programs participated in the study. Pharmacy resident-provider pairs conducted home visits for all patients in the study. Pharmacy residents assessed the subjective risk (high, medium, low) of medication nonadherence using information obtained from home visits (health literacy, support network, medications, and detection of something unexpected related to medications). An electronic medical record–based risk score was simultaneously calculated using screening tool components (i.e., electronic frailty index score, LACE+ index [length of stay in the hospital, acuity of admission, comorbidity, emergency department utilization in the 6 months before admission], and 2015 American Geriatric Society Beers Criteria). Outcome measuresThe electronic medical record–based screening tool numerical risk scores were compared with pharmacy resident subjective risk assessments using tree-based classification models to determine screening tool components that best predicted pharmacy residents’ subjective assessment of patients' likelihood of benefit from in-home pharmacist medication review. Following the study, satisfaction surveys were given to providers and pharmacy residents. ResultsThe best predictor of high-risk patients was an electronic frailty index score greater than 0.32 (indicating very frail) or LACE+ index greater than or equal to 59 (at high risk for hospital readmission). Pharmacy residents and providers agreed that homebound patients at high-risk for medication noncompliance benefited from pharmacist time and attention in home visits. ConclusionIn homebound older persons, this screening tool allowed for the identification of patients at high-risk for medication nonadherence through targeted in-home pharmacist medication reviews. Further studies are needed to validate the accuracy of this tool internally and externally.

SOPHIA: An Event-Based IoT and Machine Learning Architecture for Predictive Maintenance in Industry 4.0

Predictive Maintenance (PdM) is a prominent strategy comprising all the operational techniques and actions required to ensure machine availability and to prevent a machine-down failure. One of the main challenges of PdM is to design and develop an embedded smart system to monitor and predict the health status of the machine. In this work, we use a data-driven approach based on machine learning applied to woodworking industrial machines for a major woodworking Italian corporation. Predicted failures probabilities are calculated through tree-based classification models (Gradient Boosting, Random Forest and Extreme Gradient Boosting) and calculated as the temporal evolution of event data. This is achieved by applying temporal feature engineering techniques and training an ensemble of classification algorithms to predict Remaining Useful Lifetime (RUL) of woodworking machines. The effectiveness of the proposed method is showed by testing an independent sample of additional woodworking machines without presenting machine down. The Gradient Boosting model achieved accuracy, recall, and precision of 98.9%, 99.6%, and 99.1%. Our predictive maintenance approach deployed on a Big Data framework allows screening simultaneously multiple connected machines by learning from terabytes of log data. The target prediction provides salient information which can be adopted within the maintenance management practice.

Predicting employee attrition using tree-based models

PurposeThe purpose of this study is to develop tree-based binary classification models to predict the likelihood of employee attrition based on firm cultural and management attributes.Design/methodology/approachA data set of resumes anonymously submitted through Glassdoor’s online portal is used in tandem with public company review information to fit decision tree, random forest and gradient boosted tree models to predict the probability of an employee leaving a firm during a job transition.FindingsRandom forest and decision tree methods are found to be the strongest attrition prediction models. In addition, compensation, company culture and senior management performance play a primary role in an employee’s decision to leave a firm.Practical implicationsThis study may be used by human resources staff to better understand factors which influence employee attrition. In addition, techniques developed in this study may be applied to company-specific data sets to construct customized attrition models.Originality/valueThis study contains several novel contributions which include exploratory studies such as industry job transition percentages, distributional comparisons between factors strongly contributing to employee attrition between those who left or stayed with the firm and the first comprehensive search over binary classification models to identify which provides the strongest predictive performance of employee attrition.