Median Absolute Error Research Articles

Predictive Accuracy of Intraocular Lens Formulas Calculated by Biometers with Multiple Refractive Indices According to Axial Length

Background/Objectives: This study aimed to analyze the accuracy of the SRK/T, Haigis, Barrett Universal II (BUII), Kane, and EVO intraocular lens (IOL) formulas for normal and long axial length (AL) groups using an ARGOS (Alcon, Fort Worth, TX, USA), which uses the specific refractive indices. Methods: We performed a review of patients who underwent uneventful cataract surgery with the implantation of an Acrysof IQ SN60WF IOL (Alcon, Inc.) between January 2020 and December 2021. Biometry was obtained with the ARGOS; patients were separated into subgroups based on AL as follows: normal (22.0 ≤ AL < 26.0 mm) and long (AL ≥ 26.0 mm). Outcomes included the mean error (ME), the mean absolute error (MAE), the median absolute error (MedAE), and the proportion of eyes within ±0.25, ±0.50, ±0.75, and ±1.00 diopters (D) of the prediction error. Results: A total of 191 eyes of 191 patients were included: 162 eyes of 162 patients in the normal AL group and 29 eyes of 29 patients in the long AL group. The EVO formula was the most accurate for the normal AL group, which had the lowest MAE and MedAE. The MAE and MedAE of EVO were the lowest in the long AL group; EVO showed the highest percentage of eyes within ±0.25, ±0.75, and 1.00 D compared with other formulas. Conclusions: When using an ARGOS, the EVO formula had the lowest MAE and the highest proportion of eyes within ±1.00 D of the predicted target in both the normal and high myopia groups.

3D scanner measuring preterm infants’ head circumference and cranial volume: validation in a simulated care setting

IntroductionWeekly head circumference (HC) measurements using a measuring tape is the current standard for longitudinal brain growth monitoring of preterm infants. The MONITOR3D (M3D) 3D scanner has been developed to measure both HC and cranial volume (CrV) of preterm infants within incubators. The M3D’s usability, accuracy and precision were validated in a simulated setting in a neonatal intensive care unit (NICU).Materials and methodsDuring a simulated routine care moment, NICU nurses conducted M3D scans of a preterm doll simulating an extreme low birthweight preterm (ELBW; BW < 1,000 g) infant, followed by manual HC measurements using a measuring tape. Usability was quantified by percentage of successful HC and CrV measurements from scans. HC and CrV were calculated by marking anatomical landmarks on the 3D image. Measurements were compared to the real, ground truth (GT) values of the doll’s head, defined by an accurate medical scanner. Measurement accuracy was assessed using mean or median absolute measurement error (ME), and precision by the spread of ME, represented by the 95% interval of the ME range. ME intervals were compared with preterm weekly growth increases to assess clinical usability.ResultsRegarding usability, 56 M3D scan sessions resulted in 25 successful (44.6%) HC and CrV measurements, with incomplete 3D data being the primary cause of unsuccessful scans. Accuracy of the measuring tape for HC was 0.2 cm (proportional 0.9% of GT), and precision was 1.6 cm (6.3%). M3D’s accuracy of HC was 0.4 cm (1.5%), and precision was 0.7 cm (2.9%). For CrV, M3D’s accuracy was 8.0 mL (3.8%) and precision 22.6 mL (10.8%).ConclusionThe M3D scanner is suitable for measuring HC and CrV in ELBW infants. However, current scan success rate is too low for practical usability. The M3D’s accuracy and precision are clinically sufficient, while the precision of the current measuring tape method is inadequate for preterm infants. This makes the M3D a promising alternative for HC, offering less disturbance to the infant. In the future, the M3D technique could facilitate the creation of CrV growth reference charts for ELBW infants, enhancing the accuracy of clinical growth monitoring for preterm infants.

External evaluation of intravenous vancomycin population pharmacokinetic models in adults receiving high-flux intermittent haemodialysis.

Patients undergoing haemodialysis (HD) are at greater risk of methicillin-resistant Staphylococcus aureus infections requiring intravenous vancomycin. Close vancomycin therapeutic drug monitoring is warranted in HD patients as renal clearance is the primary elimination pathway. Clinically, population pharmacokinetics (popPK) model-informed dosing is commonly used. This study aimed to perform an external evaluation of published vancomycin popPK models developed for adults undergoing high-flux intermittent HD, and to create a dosing nomogram derived from the model that performed best. A literature review was conducted through PubMed and EMBASE to identify relevant popPK models. an external dataset was collected retrospectively from patients of 2 healthcare centres in Quebec, Canada. Selected models were implemented in NONMEM (v7.5; ICON Development Solutions). Predictive performance was assessed through prediction and simulation-based diagnostics. In total, 2386 vancomycin concentrations were collected from 274 patients and 476 antibiotic courses. Four vancomycin popPK models were selected for evaluation. None of the models demonstrated overall satisfactory or clinically acceptable predictive performance. Nonetheless, Bae et al.'s model performed best with a median prediction error of 16.25% and median absolute prediction error of 34.66%. Different predictive performance was also observed for vancomycin concentrations from samples collected during and between HD sessions. All evaluated models presented poor overall predictive performance. Further studies are required, through existing popPK model parameter re-estimation or new model development, to adequately describe vancomycin pharmacokinetics for our high-flux intermittent HD patient cohort.

Performance of Minnesota Functionals on Vibrational Frequency

ABSTRACTMolecular geometry and harmonic frequency calculations are essential in thermochemical computations, with density functional theory (DFT) being widely employed for vibrational frequency predictions due to its efficiency and accuracy. In this study, we assessed the precision of 28 Minnesota based functionals with three different basis sets using the VIBFREQ1295 dataset. Scaling factors are necessary for predicting fundamental frequencies, global scaling factors were fitted by using F38/10 and VIBFREQ1295 datasets. The superior performing functionals were then fitted based on vibrational frequency ranges to obtain frequency‐range‐specific scaling factors. We observed consistent outlier across various model chemistries in vibrational frequency predictions, with alternative scaling factors showing minimal impact on reducing outlier occurrences. Besides, large basis sets are not indispensably required for fundamental frequency predictions. M06‐L, revM06‐L, SOGGA11‐X, PW6B95‐D3(BJ), CF22D, and M06‐2X consistently exhibit excellent performance across the three basis sets. When using frequency‐range‐specific scaling factors, the root mean squard errors (RMSEs) and median absolute errors (MedAEs) of almost all the selected functionals were reduced. PW6B95‐D3(BJ), CF22D, and MN12‐SX exhibited the lowest RMSEs. Comparisons were also done for different data classifications; the dataset was classified by the elements of the molecules, vibrational frequency intervals, and vibrational modes.

Accuracy of recent intraocular lens power calculation methods in post-myopic LASIK eyes

This retrospective study compared postoperative prediction errors of recent formulas using standard- or total keratometry (K or TK) for intraocular lens (IOL) power calculation in post-myopic LASIK patients. It included 56 eyes of 56 patients who underwent uncomplicated cataract surgery, with at least 1-month follow-up at Keio University Hospital in Tokyo or Hayashi Eye Hospital in Yokohama, Japan. Prediction errors, absolute errors, and percentage of eyes with prediction errors within ± 0.25 D, ± 0.50 D, and ± 1.00 D were calculated using nine formulas: Barrett True-K, Barrett True-K TK, Haigis-L, Haigis TK, Pearl-DGS, Hoffer QST, Hoffer QST PK, EVO K, and EVO PK. Statistical comparisons utilized Friedman test, Conover’s all-pairs post-hoc, Cochran’s Q, and McNemar post-hoc testing. Root-Mean-Square Error (RMSE) was compared with heteroscedastic testing. Barrett True-K TK had the lowest median predicted refractive error (-0.01). EVO PK had the smallest median absolute error (0.20). EVO PK had the highest percentage of eyes within ± 0.25 D of the predicted value (58.9%), significantly better than Haigis-L (p = 0.047). EVO PK had the lowest mean RMSE value (0.499). The EVO PK formula yielded the most accurate IOL power calculation in post-myopic LASIK eyes, with TK/PK values enhancing accuracy.

Accuracy of 7 artificial intelligence based intraocular lens power calculation formulas in extremely long Caucasian eyes: Short title: AI-based IOL calculation in extra-long eyes

To compare 7 AI-based IOL power calculation formulas in extremely long eyes DESIGN: : Retrospective accuracy and validity analysis. SETTING: Kyiv Clinical Ophthalmology Hospital Eye Microsurgery Center, Ukraine STUDY POPULATION: : Patients with highly myopic eyes, who underwent uneventful phacoemulsification OBSERVATION PROCEDURES: Prior to cataract surgery IOL power was calculated. The power of the implanted IOL was randomly selected from the outcomes of SRK/T, Holladay 2 or Barrett Universal II. Three months after phacoemulsification, refraction was measured. Post-surgery IOL power calculations were performed utilizing the following formulas: Hill-RBF 3.0, Kane, PEARL-DGS, Ladas Super Formula AI (LSF AI), Hoffer QST, Karmona and Zhu-Lu. root mean square absolute error (RMSAE), median absolute error (MedAE) and percentage of eyes with prediction error (PE) within ±0.50 D RESULTS: : Forty eight eyes with axial length exceeding 30.00 mm were studied. Hill-RBF 3.0 yielded the lowest RMSAE (0.788) with statistical superiority only over Karmona (0.956, p=0.021). In terms of MedAE, outcomes obtained by Hoffer QST (0.442) and Hill-RBF (0.490) were statistically significant vs LSF AI (0.800, p=0.013, p=0.008, respectively). The highest percentage of eyes with PE within ±0.50 D was achieved by Hill-RBF 3.0, Kane and Hoffer QST (54.17% each) statistically significant as follows: both Hill-RBF and Kane vs LSF AI (27.08%) and Karmona (39.58%), and Hoffer QST vs LSF AI. All tested formulas demonstrated comparable trueness, with Hill-RBF 3.0 being more accurate than Karmona (RMSAE), and LSF AI being less accurate than Hoffer QST and Hill-RBF 3.0 (MedAE).

Ready-to-use Models Built Using a Diverse Set of 266 Aroma Compounds for the Estimation of Gas Chromatographic Retention Indices for the 50%-Cyanopropylphenyl-50%-Dimethylpolysiloxane Stationary Phase.

Retention index prediction based on the molecule structure is not often used in practice due to low accuracy, the need to use paid software to calculate molecular descriptors (MD), and the narrow applicability domain of many models. In recent years, relatively accurate and versatile deep learning (DL)-based models have emerged. These models are now used in practice as an additional criterion in gas chromatography-mass spectrometry identification. The DB-225ms stationary phase (usually described as 50%-cyanopropylphenyl-50%-dimethylpolysiloxane in available sources) is widely used, but ready-to-use retention index estimation models are not available for it. This study presents such models. The models are linear and use simple constitutional MD and retention indices predicted by DL for the DB-WAX and DB-624 stationary phases as MD (we show that it is their use that allows us to achieve satisfactory accuracy). The accuracy obtained for a completely unseen hold-out test set: root mean square error 73.2; mean absolute error 45.7; median absolute error 22.0. The models were trained using a retention data set of 266 volatile compounds. All calculations can be performed using the convenient open-source software CHERESHNYA. The final equations are implemented as a spreadsheet and a code snippet and are available online: https://doi.org/10.6084/m9.figshare.26800789.

Data-Driven Approaches for State-of-Charge Estimation in Battery Electric Vehicles Using Machine and Deep Learning Techniques

One of the most important functions of the battery management system (BMS) in battery electric vehicle (BEV) applications is to estimate the state of charge (SOC). In this study, several machine and deep learning techniques, such as linear regression, support vector regressors (SVRs), k-nearest neighbor, random forest, extra trees regressor, extreme gradient boosting, random forest combined with gradient boosting, artificial neural networks (ANNs), convolutional neural networks, and long short-term memory (LSTM) networks, are investigated to develop a modeling framework for SOC estimation. The purpose of this study is to improve overall battery performance by examining how BEV operation affects battery deterioration. By using dynamic response simulation of lithium battery electric vehicles (BEVs) and lithium battery packs (LIBs), the proposed research provides realistic training data, enabling more accurate prediction of SOC using data-driven methods, which will have a crucial and effective impact on the safe operation of electric vehicles. The paper evaluates the performance of machine and deep learning algorithms using various metrics, including the R2 Score, median absolute error, mean square error, mean absolute error, and max error. All the simulation tests were performed using MATLAB 2023, Anaconda platform, and COMSOL Multiphysics.

Predicting Phylogenetic Bootstrap Values via Machine Learning.

Estimating the statistical robustness of the inferred tree(s) constitutes an integral part of most phylogenetic analyses. Commonly, one computes and assigns a branch support value to each inner branch of the inferred phylogeny. The still most widely used method for calculating branch support on trees inferred under Maximum Likelihood (ML) is the Standard, non-parametric Felsenstein Bootstrap Support (SBS). Due to the high computational cost of the SBS, a plethora of methods has been developed to approximate it, for instance, via the Rapid Bootstrap (RB) algorithm. There have also been attempts to devise faster, alternative support measures, such as the SH-aLRT (Shimodaira-Hasegawa-like approximate Likelihood Ratio Test) or the UltraFast Bootstrap 2 (UFBoot2) method. Those faster alternatives exhibit some limitations, such as the need to assess model violations (UFBoot2) or unstable behavior in the low support interval range (SH-aLRT). Here, we present the Educated Bootstrap Guesser (EBG), a machine learning-based tool that predicts SBS branch support values for a given input phylogeny. EBG is on average 9.4 (σ = 5.5) times faster than UFBoot2. EBG-based SBS estimates exhibit a median absolute error of 5 when predicting SBS values between 0 and 100. Furthermore, EBG also provides uncertainty measures for all per-branch SBS predictions and thereby allows for a more rigorous and careful interpretation. EBG can, for instance, predict SBS support values on a phylogeny comprising $1654$ SARS-CoV2 genome sequences within 3 hours on a mid-class laptop. EBG is available under GNU GPL3.

Automated calculation and reporting of vancomycin area under the concentration-time curve: a simplified single-trough concentration-based equation approach.

Vancomycin, a crucial antibiotic for Gram-positive bacterial infections, requires therapeutic drug monitoring (TDM). Contemporary guidelines advocate for AUC-based monitoring; however, using Bayesian programs for AUC estimation poses challenges. We aimed to develop and evaluate a simplified AUC estimation equation using a steady-state trough concentration (Ctrough) value. Utilizing 1,034 TDM records from 580 general hospitalized patients at a university-affiliated hospital in Ulsan, we created an equation named SSTA that calculates the AUC by applying Ctrough, body weight, and single dose as input variables. External validation included 326 records from 163 patients at a university-affiliated hospital in Seoul (EWUSH) and literature data from 20 patients at a university-affiliated hospital in Bangkok (MUSI). It was compared with other AUC estimation models based on the Ctrough, including a linear regression model (LR), a sophisticated model based on the first-order equation (VancoPK), and a Bayesian model (BSCt). Evaluation metrics, such as median absolute percentage error (MdAPE) and the percentage of observations within ±20% error (P20), were calculated. External validation using the EWUSH data set showed that SSTA, LR, VancoPK, and BSCt had MdAPE values of 6.4, 10.1, 6.6, and 7.5% and P20 values of 87.1, 82.5, 87.7, and 83.4%, respectively. External validation using the MUSI data set showed that SSTA, LR, and VancoPK had MdAPEs of 5.2, 9.4, and 7.2%, and P20 of 95, 90, and 95%, respectively. Owing to its decent AUC prediction performance, simplicity, and convenience for automated calculation and reporting, SSTA could be used as an adjunctive tool for the AUC-based TDM.

Augmented and virtual reality imaging for collaborative planning of structural cardiovascular interventions: a proof-of-concept and validation study.

Virtual reality (VR) and augmented reality (AR) have led to significant advancements in cardiac preoperative planning, shaping the world in profound ways. A noticeable gap exists in the availability of a comprehensive multi-user, multi-device mixed reality application that can be used in a multidisciplinary team meeting. A multi-user, multi-device mixed reality application was developed, supporting AR and VR implementations. Technical validation involved a standardized testing protocol and comparison of AR and VR measurements regarding absolute error and time. Preclinical validation engaged experts in interventional cardiology, evaluating the clinical applicability prior to clinical validation. Clinical validation included patient-specific measurements for five patients in VR compared with standard computed tomography (CT) for preoperative planning. Questionnaires were used at all stages for subjective evaluation. Technical validation, including 106 size measurements, demonstrated an absolute median error of 0.69mm (0.25 to 1.18mm) compared with ground truth. The time to complete the entire task was on average, with VR measurements being faster than AR ( versus , ). On clinical validation of five preoperative patients, there was no statistically significant difference between paired CT and VR measurements (0.58 [95% CI, to 2.74], ). Questionnaires showcased unanimous agreement on the user-friendly nature, effectiveness, and clinical value. The mixed reality application, validated through technical, preclinical, and clinical assessments, demonstrates precision and user-friendliness. Further research of our application is needed to validate the generalizability and impact on patient outcomes.

Accuracy of smartwatches for the remote assessment of exercise capacity

Exercise capacity is a strong independent predictor of cardiovascular and all-cause mortality. The utilization of well-established submaximal tests of exercise capacity such as the 6-min walk test (6MWT), 3-min step test (3MST) and 10-chair rise test (10CRT) in the community would improve patient care but requires remote monitoring technology. Consumer grade smartwatches provide such an opportunity, however, their accuracy in measuring physiological responses to these tests is unclear. The aim of this study was to determine the accuracy of consumer grade smartwatches in assessing exercise capacity to develop a framework for remote, unsupervised testing. 16 healthy adults (7 male (44%), age median 27 [interquartile range (IQR) 26,29] years) performed 6MWTs using two protocols: (1) standard—straight 30 m laps (6MWT-standard) and 2) continuous lap—circular 240 m laps around a park (6MWT-continuous lap), 3MSTs and 10CRTs. Each one of these four tests was performed three times across two clinic visits. Each participant was fitted with a Garmin Vivoactive4 and Fitbit Sense smartwatch to measure three parameters: distance, step counts and heart rate (HR) response. Reference measures were a meter-wheel, hand tally counter and ECG, respectively. Mean HR was measured at rest, peak exercise and recovery. Agreement was measured using Bland–Altman analysis for repeated measures and summarized as median absolute percentage errors (MAPE). Distance during 6MWT-continuous lap had better agreement than during 6MWT-standard for both Garmin (MAPE: 6.4% [3.0, 10.4%] versus 20.1% [13.9, 28.4%], p < 0.001) and Fitbit (8.0% [2.9, 10.1% versus 18.8% [15.2, 28.1%], p < 0.001). Garmin measured step count more accurately than Fitbit (MAPE: 1.8% [0.9, 2.9%] versus 8.0% [2.6, 12.3%], p < 0.001). Irrespective of test, both devices showed excellent accuracy in measuring HR at rest and recovery (≤ 3%), while accuracy decreased during peak exercise (Fitbit: ~ 12% and Garmin: ~ 7%). In young adults without mobility difficulties, exercise capacity can be measured remotely using standardized tests and consumer grade smartwatches.

Satellite-derived bathymetry using Sentinel-2 in mesotidal coasts

Coastal zones are strategic environments of high socioeconomic, political, and ecological value, with over half of the world’s population residing within 200 km of the coast. This proximity highlights their vulnerability to extreme events, which are exacerbated by global changes, leading to significant coastal impacts such as erosion, flooding, and ecosystem services deterioration. Consequently, efficient and operational methodologies for continuous monitoring are urgently needed to face these challenges. Bathymetric data are essential for understanding coastal dynamics, yet traditional data collection methods are often constrained by logistical challenges and high costs. Spaceborne remote sensing techniques offer significant advantages over traditional ground-based methods, particularly in terms of cost-effectiveness and operational efficiency. Over the last half-century, different Satellite-derived bathymetry (SDB) methodologies have been developed; however, challenges still persist. In this research, we applied a robust SDB methodology to three different study sites: Cíes Islands, Baiona Bay, and Vao beach within the Ría de Vigo, Galicia (NW Spain). These areas offer diverse and complex mesotidal environments to test for the very first time the methodology's efficacy. SDB was retrieved with a median absolute error (MedAE) ranging from 0.35 m to 1.55 m for depths up to 14 m. Results with different data source were evaluated, obtaining MedAE for nautical charts ranging from 0.46 m to 1.55 m. The precision between the data sources were quite close. In addition, multi-image composite was generated using images coinciding with both low tide (LT) and high tide (HT) conditions across the three zones. The lowest MedAE values were consistently obtained in images classified as LT (0.46 m) corresponding to Vao area. The results highlight the potential of nautical charts as a reliable source of calibration data for SDB, confirm the effectiveness of multi-image and switching models to correct artifacts and turbidity, considering tidal effects, improving single image approaches, and leverage visible bands for precise depth retrieval under varying conditions.

Can input reconstruction be used to directly estimate uncertainty of a dose prediction U-Net model?

The reliable and efficient estimation of uncertainty in artificial intelligence (AI) models poses an ongoing challenge in many fields such as radiation therapy. AI models are intended to automate manual steps involved in the treatment planning workflow. We focus in this study on dose prediction models that predict an optimal dose trade-off for each new patient for a specific treatment modality. They can guide physicians in the optimization, be part of automatic treatment plan generation or support decision in treatmentindication. Most common uncertainty estimation methods are based on Bayesian approximations, like Monte Carlo dropout (MCDO) or Deep ensembling (DE). These two techniques, however, have a high inference time (i.e., require multiple inference passes) and might not work for detecting out-of-distribution (OOD) data (i.e., overlapping uncertainty estimate for in-distribution (ID) and OOD). In this study, we present a direct uncertainty estimation method and apply it for a dose prediction U-Net architecture. It can be used to flag OOD data and give information on the quality of the doseprediction. Our method consists in the addition of a branch decoding from the bottleneck which reconstructs the CT scan given as input. The input reconstruction error can be used as a surrogate of the model uncertainty. For the proof-of-concept, our method is applied to proton therapy dose prediction in head and neck cancer patients. A dataset of 60 oropharyngeal patients was used to train the network using a nested cross-validation approach with 11 folds (training: 50 patients, validation: 5 patients, test: 5 patients). For the OOD experiment, we used 10 extra patients with a different head and neck sub-location. Accuracy, time-gain, and OOD detection are analyzed for our method in this particular application and compared with the popular MCDO andDE. The additional branch did not reduce the accuracy of the dose prediction model. The median absolute error is close to zero for the target volumes and less than 1% of the dose prescription for organs at risk. Our input reconstruction method showed a higher Pearson correlation coefficient with the prediction error (0.620) than DE (0.447) and MCDO (between 0.599 and 0.612). Moreover, our method allows an easier identification of OOD (no overlap for ID and OOD data and a Z-score of 34.05). The uncertainty is estimated simultaneously to the regression task, therefore requires less time and computationalresources. This study shows that the error in the CT scan reconstruction can be used as a surrogate of the uncertainty of the model. The Pearson correlation coefficient with the dose prediction error is slightly higher than state-of-the-art techniques. OOD data can be more easily detected and the uncertainty metric is computed simultaneously to the regression task, therefore faster than MCDO or DE. The code and pretrained model are available on the gitlab repository: https://gitlab.com/ai4miro/ct-reconstruction-for-uncertainty-quatification-of-hdunet.

A multidimensional machine learning framework for LST reconstruction and climate variable analysis in forest fire occurrence

Land Surface Temperature (LST) datasets play a crucial role in understanding the complex interplay between forest fires, climate variables, and vegetation dynamics. This study is divided into two primary parts: the first part investigates the predictive performance of a machine learning framework based on CatBoost and XGBoost models in estimating LST across different land cover classes in Alberta, Canada. On the test set, for LST-Day data, CatBoost and XGBoost achieved Median Absolute Errors (MedAE) of approximately 1.434 °C and 1.425 °C, respectively. For LST-Night data, also on the test set, the MedAE values were approximately 1.186 °C for CatBoost and 1.176 °C for XGBoost. The second part explores the intricate relationships between climatic variables—LST, precipitation, and relative humidity—forest fire occurrences, and vegetation dynamics in various subregions. The findings revealed complex interactions, with high LST, reduced precipitation, and humidity associated with increased forest fire activity and subsequent changes in vegetation patterns, particularly in the Central Mixedwood, Dry Mixedwood, and Montane subregions. A notable potential association was identified between high LST, reduced precipitation and humidity, and increased forest fire activity in these areas. These climate change impacts and fire events were found to influence ecological processes, altering species composition, reducing biodiversity, and potentially disrupting ecosystem services such as carbon sequestration and nutrient cycling. These insights are crucial for informing adaptive forest management strategies aimed at understanding and mitigating the cascading effects of climate change on fire regimes and vegetation dynamics in Alberta's diverse landscapes.

Data-driven construction of maritime traffic networks for AI-based route prediction

Abstract Predicting the routes of maritime traffic can improve economic efficiency, decrease ecological impact, and improve safety at sea. Over scales that are small (few hundred meters) and large (dozens to hundreds of kilometers), vessel trajectories have successfully been predicted by deep learning and (static) network-based approaches, respectively. We present an approach for medium to large scales (few kilometers) where (a) a maritime traffic network is automatically constructed from AIS messages, and (b) vessel trajectories are predicted as most likely paths through the network. Using three regions (Stavanger, Tromsø, and Oslo), we show that the network can capture up to ∼ 90 per cent of all maritime traffic (excluding pleasure craft) with a median absolute error of ∼ 80 meters. Vessel paths are sequences of waypoints and legs (nodes and edges) and are map-matched onto the network from vessel trajectories. Once mapped, we predict future paths for two subproblems – (i) known destination, and (ii) unknown destination. We use four algorithms (Dijkstra, Markov, MOGen, GRETEL). For known destinations, we find that Dijkstra performs best. In Stavanger (Tromsø, Oslo), Dijkstra predicts 64 (42, 68) per cent of path segments correctly and keeps the median path error below 15 (33 and 55) meters. For unknown destinations, performance depends on the forecast horizon (the number of legs k to predict). For k ≤ 5, Markov is best and predicts 62 (48, 72) per cent of legs correctly. For k &gt; 5, GRETEL performs best and predicts 54 (47, 63) per cent of legs correctly. For some types of vessels, models improve by considering vessel type. For passenger vessels, models specific to them predict ∼ 10 per cent better paths with half the distance error. For tankers, paths (and distance errors) are 6 (20) per cent worse. For auxiliary vessels, path quality is unchanged, but distance error improves ∼ 36 per cent.

Detection of sedentary time and bouts using consumer-grade wrist-worn devices: a hidden semi-Markov model

BackgroundWrist-worn data from commercially available devices has potential to characterize sedentary time for research and for clinical and public health applications. We propose a model that utilizes heart rate in addition to step count data to estimate the proportion of time spent being sedentary and the usual length of sedentary bouts.MethodsWe developed and trained two Hidden semi-Markov models, STEPHEN (STEP and Heart ENcoder) and STEPCODE (STEP enCODEr; a steps-only based model) using consumer-grade Fitbit device data from participants under free living conditions, and validated model performance using two external datasets. We used the median absolute percentage error (MDAPE) to measure the accuracy of the proposed models against research-grade activPAL device data as the referent. Bland-Altman plots summarized the individual-level agreement with activPAL.ResultsIn OPTIMISE cohort, STEPHEN’s estimates of the proportion of time spent sedentary had significantly (p < 0.001) better accuracy (MDAPE [IQR] = 0.15 [0.06–0.25] vs. 0.23 [0.13–0.53)]) and agreement (Bias Mean [SD]=-0.03[0.11] vs. 0.14 [0.11]) than the proprietary software, estimated the usual sedentary bout duration more accurately (MDAPE[IQR] = 0.11[0.06–0.26] vs. 0.42[0.32–0.48]), and had better agreement (Bias Mean [SD] = 3.91[5.67] minutes vs. -11.93[5.07] minutes). With the ALLO-Active dataset, STEPHEN and STEPCODE did not improve the estimation of proportion of time spent sedentary, but STEPHEN estimated usual sedentary bout duration more accurately than the proprietary software (MDAPE[IQR] = 0.19[0.03–0.25] vs. 0.36[0.15–0.48]) and had smaller bias (Bias Mean[SD] = 0.70[8.89] minutes vs. -11.35[9.17] minutes).ConclusionsSTEPHEN can characterize the proportion of time spent being sedentary and usual sedentary bout length. The methodology is available as an open access R package available from https://github.com/limfuxing/stephen/. The package includes trained models, but users have the flexibility to train their own models.

Analyzing the impact of investor sentiment on S&amp;P 500 prices using deep learning models

Purpose This study aims to investigate to what extent the predictability of the standard and poor’s 500 (S&amp;P 500) price levels is enhanced by investors’ sentiments extracted from social media content, specifically platform X. Design/methodology/approach Two recurrent neural network (RNN) models are developed. The first RNN model is merely based on historical records and technical indicators. In addition to the variables included in the first RNN model, the second RNN model comprises the outputs of the sentiment analysis, performed using the TextBlob library. The study was conducted between December 28, 2011, and December 30, 2021, over 10 years, to obtain better results by feeding the RNN models with a significant quantity of data by extending the period and capturing an extensive timespan. Findings Comparing the performance of both models reveals that the second model, with sentiment analysis inputs, yields superior outcomes. The mean absolute error (MAE) of the second model registered 72.44, approximately 50% lower than the MAE of the technical model, its percentage value, the mean absolute percentage error, recorded 2.16%, and finally, the median absolute percentage error reached a value of 1.30%. This underscores the significant influence of digital platforms in influencing the behavior of certain assets like the S&amp;P 500, emphasizing the relevance of sentiment analysis from social media in financial forecasting. Originality/value This study contributes to the growing body of literature by highlighting the enhanced predictive power of deep learning models that incorporate investor sentiment from social media, thereby advancing the application of behavioral finance in financial forecasting.

Accuracy of an internationally validated genetic-guided warfarin dosing algorithm compared to a clinical algorithm in an Arab population

PurposeTo identify the impact of CYP2C9*2, *3, VKORC1−1639 G>A and CYP4F2*3 on warfarin dose in an Arab population. To compare the accuracy of a clinical warfarin dosing (CWD) versus genetic warfarin dosing algorithms (GWD) during warfarin initiation. MethodsA cohort of Arab patients newly starting on warfarin had their dose calculated using CWD published in www.warfarindosing.org and were followed for 1 month. Each patient provided a saliva sample. DNA was extracted, purified and genotyped for VKORC−1639 G>A, CYP2C9*2, CYP2C9*3 and CYP4F2*3. After reaching warfarin maintenance dose, the dose was recalculated using the GWD and median absolute error (MAE) and the percentage of warfarin doses within 20% of the actual dose were calculated and compared for the two algorithms. ResultsThe study enrolled 130 patients from 12 Arabian countries. Compared to those with wild type, carriers of reduced function alleles in CYP2C9 required significantly lower median (IQR) warfarin weekly dose [24.5 (15.3) vs. 35 (29.8) mg/week, p=0.006]. With regards to VKORC, patients with AA genotype had a significantly lower median (IQR) weekly warfarin dose compared to AG and GG [21(10.5) vs 29.4 (21), p<0.001 for AA vs AG, p<0.001 for AA vs GG]. The MAE (IQR) for the weekly dose of the GWD was significantly lower compared to CWD [8.1 (10.5) vs 12.4 (12.6) (p<0.001)]. ConclusionCYP2C9 and VKORC1 variants are important determinants of warfarin dose in the Arab population. The use of the genetic and clinical factors led to better warfarin dose estimation when compared to clinical factors alone.

Extracting seizure control metrics from clinic notes of patients with epilepsy: A natural language processing approach

ObjectivesMonitoring seizure control metrics is key to clinical care of patients with epilepsy. Manually abstracting these metrics from unstructured text in electronic health records (EHR) is laborious. We aimed to abstract the date of last seizure and seizure frequency from clinical notes of patients with epilepsy using natural language processing (NLP). MethodsWe extracted seizure control metrics from notes of patients seen in epilepsy clinics from two hospitals in Boston. Extraction was performed with the pretrained model RoBERTa_for_seizureFrequency_QA, for both date of last seizure and seizure frequency, combined with regular expressions. We designed the algorithm to categorize the timing of last seizure (“today”, “1–6 days ago”, “1–4 weeks ago”, “more than 1–3 months ago”, “more than 3–6 months ago”, “more than 6–12 months ago”, “more than 1–2 years ago”, “more than 2 years ago”) and seizure frequency (“innumerable”, “multiple”, “daily”, “weekly”, “monthly”, “once per year”, “less than once per year”). Our ground truth consisted of structured questionnaires filled out by physicians. Model performance was measured using the areas under the receiving operating characteristic curve (AUROC) and precision recall curve (AUPRC) for categorical labels, and median absolute error (MAE) for ordinal labels, with 95 % confidence intervals (CI) estimated via bootstrapping. ResultsOur cohort included 1773 adult patients with a total of 5658 visits with reported seizure control metrics, seen in epilepsy clinics between December 2018 and May 2022. The cohort average age was 42 years old, the majority were female (57 %), White (81 %) and non-Hispanic (85 %). The models achieved an MAE (95 % CI) for date of last seizure of 4 (4.00–4.86) weeks, and for seizure frequency of 0.02 (0.02–0.02) seizures per day. ConclusionsOur NLP approach demonstrates that the extraction of seizure control metrics from EHR is feasible allowing for large-scale EHR research.