Simple Regression Model Research Articles

ANALYSIS OF THE EFFECTS OF SELECTED QUANTITIES ON THE FOULING OF NATURAL GAS COOLER PIPES

In this article, the effects of selected quantities on the thickness of a fouling layer that is formed on the internal surface of the pipes of natural gas (NG) coolers are discussed. Models created by applying an analytical method, a dimensional analysis, and a multiple linear regression were compared. The most complex model was the analytical model with a total of 16 relevant physical quantities. From the practicality point of view, that model was very complicated since it required applying the iterative process. The dimensional model comprised 6 similarity criteria. The difference between the fouling layer thicknesses identified analytically and those identified using a model based on a dimensional analysis did not exceed 3.06 %. Out of the total number of relevant quantities, only those that had the strongest effects on the fouling layer thickness were selected with the use of MinitabX software. Three regression models, containing 6, 5, and 4 variables, were thus created. The difference between the results obtained from the simplest regression model and the values obtained from the dimensional model amounted to a maximum of 3.20 %. The use of the presented regression models in technical practice may be recommended.

Consumer Credit: a Historical Review of the Impact on the Ecuadorian Economy, Period 2005 – 2024

Objective: Analyze consumer credit and its impact on the Ecuadorian economy, period 2005 – 2024. Theoretical Framework: The theoretical framework of the Keynesian school was used, in which the role of financial institutions is crucial for driving economic growth through the financing of consumption. Method: Quantitative research with a depth of descriptive analysis; 78 quarterly data points were used for Pearson's tests and for the simple time series regression model. Results and Discussion: The trend of consumer credit was positive during the research period; however, its economic contribution fluctuated according to the economic context or cycle. It was significant during times of boom, when it accounted for approximately 4.38% of PIB between 2011 and 2014. In contrast, during the 2005–2024 period, the results showed that it had a moderate impact on PIB, but it was not significant for economic growth, as the Ecuadorian economy operates in an imperfect competition market. Originality/Value: This study allows for the verification of the theoretical assumptions in the Ecuadorian case; additionally, it provides relevant conclusions that explain the model and characteristics of the market.

Towards Sustainable Material Design: A Comparative Analysis of Latent Space Representations in AI Models

In this study, we employed machine learning techniques to improve sustainable materials design by examining how various latent space representations affect the AI performance in property predictions. We compared three fingerprinting methodologies: (a) neural networks trained on specific properties, (b) encoder–decoder architectures, and c) traditional Morgan fingerprints. Their encoding quality was quantitatively compared by using these fingerprints as inputs for a simple regression model (Random Forest) to predict glass transition temperatures (Tg), a critical parameter in determining material performance. We found that the task-specific neural networks achieved the highest accuracy, with a mean absolute percentage error (MAPE) of 10% and an R2 of 0.9, significantly outperforming encoder–decoder models (MAPE: 19%, R2: 0.76) and Morgan fingerprints (MAPE: 24%, R2: 0.6). In addition, we used dimensionality reduction techniques, such as principal component analysis (PCA) and t-distributed stochastic neighbour embedding (t-SNE), to gain insights on the models’ abilities to learn relevant molecular features to Tg. By offering a more profound understanding of how chemical structures influence AI-based property predictions, this approach enables the efficient identification of high-performing materials in applications that range from water decontamination to polymer recyclability with minimum experimental effort, promoting a circular economy in materials science.

Molecular Design for Cardiac Cell Differentiation Using a Small Data Set and Decorated Shape Features.

The discovery of small organic compounds for inducing stem cell differentiation is a time- and resource-intensive process. While data science could, in principle, streamline the discovery of these compounds, novel approaches are required due to the difficulty of acquiring training data from large numbers of example compounds. In this paper, we present the design of a new compound for inducing cardiomyocyte differentiation using simple regression models trained with a data set containing only 80 examples. We introduce decorated shape descriptors, an information-rich molecular feature representation that integrates both molecular shape and hydrophilicity information. These models demonstrate improved performance compared to ones using standard molecular descriptors based on shape alone. Model overtraining is diagnosed using a new type of sensitivity analysis. Our new compound is designed using a conservative molecular design strategy, and its effectiveness is confirmed through expression profiles of cardiomyocyte-related marker genes using real-time polymerase chain reaction experiments on human iPS cell lines. This work demonstrates a viable data-driven strategy for designing new compounds for stem cell differentiation protocols and will be useful in situations where training data is limited.

Analysis of different populations accessing online overdose response training and harm reduction supplies (ADORES)

BackgroundAn online mail order naloxone and harm reduction supply program was created by an interdisciplinary team at the University of Rhode Island College of Pharmacy and hosted on the university website (UNIV). The program was subsequently funded by the Rhode Island Department of Health (DOH) and added to the DOH website. This study compares demographic characteristics of the two populations submitting requests through the program’s distinct access points, UNIV and DOH, to those of populations at-risk for overdose.MethodsThis is a retrospective comparative analysis using voluntarily provided information from mail order request forms submitted through two websites from June 2020 through October 2023. The primary objective was to compare demographic characteristics of requesters through the two access points to those of individuals at-risk for experiencing or responding to an overdose. Descriptive statistics were used to characterize the two distinct populations. Chi-square tests were performed to determine if statistically significant differences in population demographics existed between access points. Odds ratios were estimated using a simple logistic regression model to assess the relationship between access point and demographic characteristic to determine if either access point had a greater likelihood of reaching individuals with at-risk demographics.ResultsA total of 5,783 (UNIV = 1662; DOH = 4121) mail order request forms were included in the analysis. Participants who completed requests through UNIV were more likely to be 44 years old or younger and/or reside in rural areas. Participants who submitted requests through DOH were more likely to reside in non-rural and/or low-income areas, and/or identify as gender minorities. Additionally, UNIV respondents were more likely to be first-time naloxone requesters while DOH respondents were more likely to have obtained and used naloxone before.ConclusionsResults establish that different access points can reach different at-risk population subsets and support the use of multiple access points and advertising strategies to expand the reach of online mail order harm reduction programs.Study registrationThe study was retrospectively approved by the University of Rhode Island Institutional Review Board (IRB reference #2124391-2).

Application of neural networks for determining the parameters of PVT-models used in solving hydrodynamic modeling problems

The article presents an overview of the application of machine learning models for problems of modeling the phase state of hydrocarbon systems. Determination of the state of a hydrocarbon system is possible in the presence of nonhydrocarbon components when performing hydrodynamic calculations on multiphase models. Such calculations are in demand when solving problems of both forecasting and adapting models to actual development data. It is shown that when adjusting models to actual data, the accuracy of reproducing laboratory experiments is in some cases higher than when using equations of state. But in some cases, the use of neural networks turns out to be useless. For relatively “simple” fluid systems, an increase in the complexity of the tool used does not provide an increase in accuracy compared to simple regression models. Another problem is the lack of a working tool that would completely replace the use of equations of state to assess the state of a fluid model when performing hydrodynamic calculations.

Fetal Gastroschisis: Maternal Epidemiologic Profile and Influence of Possible Risk Factors on Neonatal Outcomes.

The aim of this study was to evaluate the maternal epidemiological profile for gastroschisis as well as to assess whether the presence of meconium at birth was associated with adverse neonatal outcomes. This was a retrospective observational cohort study that included 52 neonates with gastroschisis whose mothers had delivery at our service. The variables observed were the following: 1) socio-demographic data, 2) ultrasonographic data, 3) birth data, and 4) neonatal outcomes. Simple (univariate) and multiple (multivariate) logistic regression models were used to examine the influence of meconium on neonatal outcomes. Mean maternal age and body mass index (BMI) of the pregnant women were 21.9 years and 23.3 kg/m2, respectively. There was a similarity between mixed (50%) and white (48.1%) races, and 27.7% (13) reported to use an oral contraceptive (OC). Mean birth weight was 2,386.9 grams, and 24 of the neonates (46.2%) were premature (<37 weeks). The majority of neonates had neonatal sepsis (40.4%) and 39 were admitted to the neonatal intensive care unit with a median length of stay of 31 days. Forty neonates required mechanical ventilation, with a median duration of 5 days. Primary closure was achieved in 50% (26) of cases. Although not significant and with low precision, the point estimate suggested that the presence of meconium at birth, on average, reduced the chance of primary closure (OR=0.26, CI 95% [0.06; 1.16], p=0.077). Pregnant women with gastroschisis fetuses were more likely to be young, primigravida, with a normal BMI, white or mixed, and with a history of OC use. There was no evidence of an association between the presence of meconium at birth and adverse perinatal outcomes.

COMPARATIVE ANALYSIS OF REGRESSION MODELS FOR STOCK PRICE PREDICTION: LINEAR, SUPPORT VECTOR, POLYNOMIAL, AND LASSO

This paper investigates the performance of various regression-based machine learning techniques for short-term stock price prediction. Four regression algorithms – linear regression, support vector regression (SVR), polynomial regression and LASSO regression – were applied to Apple Inc’s historical price data for two-years ending on October 1st, 2024, to predict the next day’s closing stock price. Technical indicators such as moving average, relative strength index (RSI) and lag features were included in the regressions to improve prediction performance. Evaluation metrics such as Mean Absolute Error (MAE), Mean Squared Error (MSE), and R-squared (R²) were employed to assess the models’ performance. Results show that linear regression and LASSO regression were the most performant models, reaching R² values of more than 0.95, with minimal error values. While SVR yielded the poorest results when accounting for all metrics. Overall, the study highlights the predictive power of simpler regression models over more complex ones in stock price predictions and offers recommendations for model selection.

Alveolar Bone Reconstruction Simultaneous to Implant Removal due to Advanced Peri-Implantitis Defects: A Proof of Concept.

To evaluate the safety and effectiveness of alveolar bone reconstruction simultaneous to implant removal due to peri-implantitis. Partial or fully dentulous patients subjected to implant removal due to advanced peri-implantitis (≥ 50% of bone loss) lesions and seeking to have the failed implant replaced for esthetic or functional reasons were consecutively included. Guided bone regeneration was performed by means of a mixture of xenograft and autogenous bone and a ribose cross-linked barrier membrane. Re-entry for implant placement was performed at 4-month follow-up. Overall, six radiographic variables were assessed before (T0) and after (T1) alveolar bone reconstruction at four levels in ridge width (RW) and height (RH). Peri-implant conditions were evaluated at latest follow-up. Simple and multiple binary logistic regression models were calculated using generalized estimation equations to evaluate the effect of baseline upon reconstructive outcomes. In total, 20 patients (nimplant = 39) met the inclusion criteria. Alveolar RW and RH were augmented from T0 to T1 at all levels. All implants achieved primary stability. Only ~13% were subjected to ancillary bone regeneration simultaneous to implant placement. After a mean follow-up period after loading of ~2.2 years, ~70% implants demonstrated peri-implant health, while mucositis was diagnosed in the remaining implants. The performance of alveolar bone reconstruction in residual partially contained defects simultaneous to implant removal due to peri-implantitis lesions demonstrates being safe and effective for implant site development.

Effects of screening coverage and screening quality assurance on cervical cancer mortality: Implication for integrated framework to monitor global implementation of cervical cancer screening programmes.

Cervical cancer is a global health threat and a manifestation of inequality, and screening is an effective intervention. However, little is known about how screening coverage and quality assurance, influence cervical cancer mortality. We aimed to investigate the association between screening coverage, screening quality assurance and cervical cancer mortality among women from countries at different developmental levels. We obtained data on age-standardised mortality from cervical cancer from the GLOBOCAN 2020 database; coverage of cervical cancer screening from World Health Organization (WHO) Global Health Observatory; and cervical screening programme settings and quality assurance from the Cancer Screening in Five Continents (CanScreen5) database. We assessed the dependency of cervical cancer age-standardised mortality on screening coverage and quality assurance by simple and multiple regression models. We also used linear regression models to identify factors that improved the screening coverage. The study included data from 53 countries. Reduced mortality was associated with increased screening programme quality assurance in 22 high-development countries. In 31 low-development countries, screening coverage in women aged 30-49 years was inversely associated with cervical cancer mortality. Political commitment (documentation of the cervical cancer screening policy as law) and financial support (treatment services provided free of charge) positively associated with screening coverage. Screening programmes need strengthening commensurate with local resources and context. Priority should be given to improving screening coverage through stronger political commitment and financial support in low-development countries, and to ensuring good performance at all levels in high-development countries.

The refugee and migrant health "global competency standards for health workers": results of a survey in general practitioner trainees in Sicily.

Refugees and migrants may represent the most vulnerable communities in many societies. Health systems should be sensitive to needs of refugees and migrants. The document "The Refugee and Migrant Health: Global Competency Standards for Health Workers (the Standards)" identifies the competencies and areas of practice of health workers working with refugees and migrants. The aim of this study is to provide an analysis of these competences and training needs, identifying the educational priorities for the implementation of the Standards in Sicily, Italy. A cross-sectional analytical pilot study was conducted by administering a questionnaire, in electronic format "Google Form ®," to doctors attending the Medical Training Course for General Practitioners in Sicily (Italy). Data obtained were collected in a Microsoft Excel database and analyzed with IBM SPSS Software 24 version. Absolute and relative frequencies were calculated for relevant categorical variables. Univariate analysis of the selected variables was subsequently carried out. The significance level chosen was a p value <0.05 (two-tailed). A total of 192 General Practitioner (GP) trainees responded to the questionnaire. They were asked if their training course comprised a Global Health Course that included the topic of health protection and social and health care for migrant populations and the 65.4% of them answered "No" or "Do not Know." GP trainees were also asked if they considered useful to include a Global Health Course dedicated to the management of patients with a migratory background within the training course in general medicine. Of the total 192 participants, 81.9% answered "Yes." Overall, in a simple regression model, the perception of having addressed migrants' health needs is positively correlated with having attended a Global Health Course (OR = 3.34 95%CI 1.2-9.1; p = 0.018). This study identified educational priorities for the implementation of the Standards in Sicily for doctors attending the Medical Training Course for General Practitioners. We hope that the results of this study will guide and inform possible future projects to implement the Standards at a national level.

Bayesian updating of relationships between crack width and corrosion level in reinforced concrete based on a large set of experimental data

AbstractThe evaluation of rebar corrosion in reinforced concrete (RC) structures presents a significant challenge for structural engineers, with on‐site visual inspections and crack width measurements playing a key role in the preliminary assessment. A general approach is to use corrosion models to predict the corrosion level of existing structures, and subsequently update these predictions by means of on‐site data. In this process, empirical models are often applied that relate the observed damage, such as corrosion‐induced cracks, to the corrosion level. However, these models have large uncertainties and are often not applicable to conditions that deviate from the test setups used to derive the empirical relations. This research aims to expand the applicability of these practical, existing empirical models for on‐site corrosion assessment through Bayesian updating techniques. To this aim, a Bayesian framework is developed to update crack width–corrosion models. The Bayesian updating methodology is illustrated for a simple regression model, and for a more complex relation that accounts for additional variables. A novel online database (KUL‐edCCRC) is used that is published accompanying this paper. The obtained results illustrate that the updated models have better agreement with the experimental results, and it is found that the confidence intervals of the regression parameters decrease for an increasing number of observations, even for a small number of additional datapoints. The results hence prove the efficiency of the approach to integrate information from new observations with prior information to adjust the crack width–corrosion model for specific conditions or cases, resulting in a more relevant model as more information becomes available.

Subaerial Profiles at Two Beaches: Equilibrium and Machine Learning

AbstractShoreline position (e.g., beach width) is a critical component of flooding and overtopping forecasts but difficult to predict accurately. We model beach width changes with a supervised machine learning (ML) approach informed by equilibrium principles. The time history of wave energy anomalies that force equilibrium models is used as an ML input feature. The sweeping simplifying equilibrium model assumptions relating beach width change to anomalies are replaced with data‐based ML results. Supervised learning regression methods including linear, support vector, decision trees, and ensemble regressors are tested. Observations for model training and testing includes weekly to quarterly beach elevation surveys spanning approximately 500 m alongshore and 8 years at two beaches, each supplemented with several months of ∼100 sub‐weekly surveys. These beaches, with different sediment types (sand vs. sand‐cobble mix), both widen in summer in response to the seasonal wave climate, in agreement with a generic equilibrium model. Differences in backshore erodability contribute to differing beach responses in the stormiest (El Niño) year that are reproduced by a simple extra trees regression model but not by the equilibrium model. With sufficiently extensive training data, the ML model outperforms equilibrium by providing flexibility and complexity in the response to wave forcing. The present ML and equilibrium models both fail to simulate a uniquely stunted beach recovery unlike other recoveries in the training data.

Positive body image and childbearing intention: A cross-sectional study

BackgroundChildbearing plays a crucial role in population movements and one of the most important factors in societal development. Nowadays, given the decrease in fertility rates, especially in developed and some developing countries, and the multitude of factors affecting these rates, this study aims to investigate the association of positive body image and childbearing intention in Iran. MethodsThis descriptive and cross-sectional study involved all health care centers of Qaen city, eastern Iran. The list of reproductive-age women and men in each center was extracted from the Ministry of Health of Iran's database. Ultimately, 494 individuals were included in the study using available convenience sampling method. Data were collected applying demographic information questionnaire, “body appreciation scale-2 (BAS-2) and childbearing intention questionnaire. Statistical analysis was performed using SPSS.26 software. ResultsOf 494 participants in this study, 50 % were men. The average age of the participants was 32.25 years (SD = 8.02). The average age of women was higher than that of men. In this study, the mean and SDs of positive body image score and childbearing intention were 39.49 (SD = 8.14) and 2.36 (SD = 1.63), respectively. In general, a positive and significant correlation was found between age (r = 0.126, p = 0.005), income (r = 0.137, p = 0.002), physical activity (r = 0.127, p = 0.005) and positive body image (r = 0.482, p < 0.001) with childbearing intention. After categorizing childbearing intention into two groups, “less than 2 children" and “2 children and more", the results of simple regression modeling showed that age, income, physical activity, body appreciation, bachelor's education and being an employee are associated with intention to have two or more children. Using multiple logistic regression modeling, a statistically significant association was observed between the childbearing intention at an older age, (OR = 1.07, p < 0.001) as well as positive body image (OR = 1.20, p < 0.001). ConclusionsThe findings indicated a significant and positive association between positive body image and couples' fertility intention. In other words, alongside external factors, internal factors such as body appreciation can influence childbearing intention and population growth rate, which highlights the importance of considering internal factors in childbearing intention.

Regression-based model predicting cost contingencies for road network projects

Cost contingencies represent a significant markup percentage for construction projects. Many construction projects experience cost overruns due to unexpected events before or during the construction phase. Determining the cost contingency for construction projects is one of the most frequent and widespread financial problems many countries face. Since the amount included in the estimates is always insufficient to absorb the actual cost overruns of the project. The main objective of this paper is to develop a simple realistic regression model for predicting the appropriate contingency amount for road network projects before the bidding stage. The regression model has been developed based on data from real case studies for 75 road network projects to adequately consider all potential risks may face the project. Good-of-fit tests were used to select the best-fitting statistical distribution for a given dataset. Based on the analysis of selected data, the log-logistic probability distribution function is the best-fit curve representing cost overrun behavior, and the cumulative distribution function has been used to determine the realistic probability of incurred cost overrun. Based on the analysis of real case studies, it is suggested that an average of 59.722% be used as a cost overrun contingency amount for road network projects. The study outcomes will help practitioners and researchers improve future accuracy/calculations of contingency amount for road construction projects and improve the decision-making process.

Lifestyle aspects are associated with common mental disorders in women over 40 years older in a population-based study

Background & AimsCommon mental disorders (CMD) are more prevalent in women, as well as noncommunicable diseases. Diet and physical activity are lifestyle modifiable factors that might help on managing these conditions. MethodsThis study aimed to investigate the association between lifestyle aspects (diet and physical activity) and common mental disorders in women aged 40+ years. Methods: This is a cross-sectional, population-based study (2015 ISA-Nutrition) with a representative sample of São Paulo-SP-Brazil urban residents. The present study was conducted with data from 467 women aged 40+ years old. Common Mental Disorders (dependent variable) were investigated by the Self-Reporting Questionnaire-20 (SRQ-20); the independent variables of interest were: the inflammatory potential of the diet was evaluated by dietary inflammatory index (DII), and physical activity level (evaluated by IPAQ) adopting the leisure dimension of physical activity. We considered as adjusting variables the presence of self-reported Noncommunicable Diseases (NCD), the age intervals (defined as 40-45, 46-50, 51-55, 56-60, 61+ years old), schooling, (according to years of formal education), Body Mass Index (BMI); and ethnicity (self-declared skin color). Simple and multi-adjusted logistic regression models were performed to investigate the associations. ConclusionsThe main findings indicate that the prevalence of CMD was 32.2%, and the frequency of having one or more NCD was 67.2%. In the final regression model, CMD was associated with the highest tertile of the DII (OR=2.215; p=0.003) and having three, and four or more NCD (OR= 6. 735; p<0.001, and OR= 3.874; p=0.033, respectively). Altogether, our results indicate that dietary inflammatory characteristics, and physical activity, along with NCD, are associated with CMD, in women aged 40+ years old, in different dimensions.

Parathyroid gland identification and angiography classification using simple machine learning methods.

Near-infrared indocyanine green angiography allows experienced surgeons to reliably evaluate parathyroid gland vitality during thyroid and parathyroid operations in order to predict postoperative function. To facilitate equal performance between surgeons, we developed an automatic computational quantification method using computer vision that portrays expert interpretation of visualized parathyroid gland near-infrared indocyanine green angiographic fluorescence signals. Near-infrared indocyanine green-parathyroid gland angiography video recordings (Fluobeam® LX, Fluoptics, Grenoble-part of Getinge-Göteborg) from patients undergoing endocrine cervical surgery in a high-volume unit were used for model development. Computation (MATLAB, Mathworks, Ireland) included segmentation-identification of the parathyroid gland (by autofluorescence), image stabilization (by linear translation) and adjusted time-fluorescence intensity profile generation. Relative upslope and maximum intensity ratios then trained a simple logistic regression model based on expert interpretation and outcome (including hypoparathyroidism), with subsequent unseen testing for validation. The model was trained on 37 patient videos (45 glands, 29 judged well perfused by parathyroid gland angiography experts), achieving feature data separation with 100% accuracy, and tested on 22 unseen videos (27 glands, 15 judged well perfused), including four in real time. Segmentation-guided parathyroid gland detection correctly identified all parathyroid glands during unseen testing along with three additional non-parathyroid gland regions (90% positive predictive value). Subsequent time-fluorescence intensity profile extraction with vitality prediction was shown feasible in all cases within 5 min, with a 96.3% model accuracy (sensitivity and specificity were 93.3 and 100% respectively) when compared with expert judgement. Automatic parathyroid gland perfusion quantification using simple machine learning computational methods discriminates parathyroid gland perfusion in concordance with expert surgeon interpretation, providing a means for near-infrared indocyanine green-parathyroid gland signal evaluation.

Incidence and risk factors of cognitive impairment in COVID-19 survivors within the first six months and its association with functional outcome

Background &amp; Objective: There is growing evidence of cognitive decline post-COVID-19, even in mild infection. We aim to evaluate the incidence of cognitive impairment, domains affected, its risk factors and the effect on function during the subacute period when rehabilitation is crucial. Methods: In this study, the incidence of impaired cognition was assessed in patients at 3 and 6 months post-COVID-19 infection between August 2021 and July 2022 at University Malaya Medical Center, with Montreal Cognitive Assessment (MoCA). The most common cognitive domains affected were identified with descriptive analysis. Associated sociodemographic and clinical factors were analyzed with simple and multiple logistic regression models. The post-COVID-19 Functional Scale (PCFS) was used to assess functional status. The correlation between cognition (MoCA score) and functional status (PCFS scale) was performed using Spearman correlation test. Results: We recruited 100 patients and found that 44% had impaired cognition at three months and 43% at six months. Patients with secondary education level (p =0.001, OR 13.541), oxygen therapy (p=0.039, OR 7.811), and obesity (p=0.029, OR 4.764) were associated with a higher risk of impaired cognition. The most affected MoCA domains were language, executive function, attention and memory. Lower MoCA score was correlated with higher PCFS grade (lower functional status) (p &lt;0.001, ρ -0.729). Conclusion: Post-COVID-19 cognitive impairments were common up to 6 months of illness and affect function. Clinicians are advised to perform cognitive screening especially in higher risk patients and provide necessary interventions.

Enhancing personalized gene expression prediction from DNA sequences using genomic foundation models

Artificial intelligence/deep learning (AI/DL) models that predict molecular phenotypes like gene expression directly from DNA sequences have recently emerged. While these models have proven effective at capturing the variation across genes, their ability to explain inter-individual differences has been limited. We hypothesize that the performance gap can be narrowed through the use of pre-trained embeddings from the Nucleotide Transformer, a large foundation model trained on 3,000+ genomes. We train a transformer model using the pre-trained embeddings and compare its predictive performance to Enformer, the current state-of-the-art model, using genotype and expression data from 290 individuals. Our model significantly outperforms Enformer in terms of correlation across individuals and narrows the performance gap with an elastic net regression approach using just the genetic variants as predictors. Although simple regression models have their advantages in personalized prediction tasks, DL approaches based on foundation models pre-trained on diverse genomes have unique strengths in flexibility and interpretability. With further methodological and computational improvements with more training data, these models may someday predict molecular phenotypes from DNA sequences with accuracy surpassing that of regression-based approaches. Our work demonstrates the potential for large pre-trained AI/DL models to advance functional genomics.

Evaluation of a Natural Language Processing Approach to Identify Diagnostic Errors and Analysis of Safety Learning System Case Review Data: Retrospective Cohort Study.

Diagnostic errors are an underappreciated cause of preventable mortality in hospitals and pose a risk for severe patient harm and increase hospital length of stay. This study aims to explore the potential of machine learning and natural language processing techniques in improving diagnostic safety surveillance. We conducted a rigorous evaluation of the feasibility and potential to use electronic health records clinical notes and existing case review data. Safety Learning System case review data from 1 large health system composed of 10 hospitals in the mid-Atlantic region of the United States from February 2016 to September 2021 were analyzed. The case review outcome included opportunities for improvement including diagnostic opportunities for improvement. To supplement case review data, electronic health record clinical notes were extracted and analyzed. A simple logistic regression model along with 3 forms of logistic regression models (ie, Least Absolute Shrinkage and Selection Operator, Ridge, and Elastic Net) with regularization functions was trained on this data to compare classification performances in classifying patients who experienced diagnostic errors during hospitalization. Further, statistical tests were conducted to find significant differences between female and male patients who experienced diagnostic errors. In total, 126 (7.4%) patients (of 1704) had been identified by case reviewers as having experienced at least 1 diagnostic error. Patients who had experienced diagnostic error were grouped by sex: 59 (7.1%) of the 830 women and 67 (7.7%) of the 874 men. Among the patients who experienced a diagnostic error, female patients were older (median 72, IQR 66-80 vs median 67, IQR 57-76; P=.02), had higher rates of being admitted through general or internal medicine (69.5% vs 47.8%; P=.01), lower rates of cardiovascular-related admitted diagnosis (11.9% vs 28.4%; P=.02), and lower rates of being admitted through neurology department (2.3% vs 13.4%; P=.04). The Ridge model achieved the highest area under the receiver operating characteristic curve (0.885), specificity (0.797), positive predictive value (PPV; 0.24), and F1-score (0.369) in classifying patients who were at higher risk of diagnostic errors among hospitalized patients. Our findings demonstrate that natural language processing can be a potential solution to more effectively identifying and selecting potential diagnostic error cases for review and therefore reducing the case review burden.