Teaching old docks new tricks with machine learning enhanced ensemble docking.

We here introduce Ensemble Optimizer (EnOpt), a machine-learning tool to improve the accuracy and interpretability of ensemble virtual screening (VS). Ensemble VS is an established method for predicting protein/small-molecule (ligand) binding. Unlike traditional VS, which focuses on a single protein conformation, ensemble VS better accounts for protein flexibility by predicting binding to multiple protein conformations. Each compound is thus associated with a spectrum of scores (one score per protein conformation) rather than a single score. To effectively rank and prioritize the molecules for further evaluation (including experimental testing), researchers must select which protein conformations to consider and how best to map each compound's spectrum of scores to a single value, decisions that are system-specific. EnOpt uses machine learning to address these challenges. We perform benchmark VS to show that for many systems, EnOpt ranking distinguishes active compounds from inactive or decoy molecules more effectively than traditional ensemble VS methods. To encourage broad adoption, we release EnOpt free of charge under the terms of the MIT license.

Predicting Basketball Shot Outcome From Visuomotor Control Data Using Explainable Machine Learning.

Quiet eye (QE), the visual fixation on a target before initiation of a critical action, is associated with improved performance. While QE is trainable, it is unclear whether QE can directly predict performance, which has implications for training interventions. This study predicted basketball shot outcome (make or miss) from visuomotor control variables using a decision tree classification approach. Twelve basketball athletes completed 200 shots from six on-court locations while wearing mobile eye-tracking glasses. Training and testing data sets were used for modeling eight predictors (shot location, arm extension time, and absolute and relative QE onset, offset, and duration) via standard and conditional inference decision trees and random forests. On average, the trees predicted over 66% of makes and over 50% of misses. The main predictor, relative QE duration, indicated success for durations over 18.4% (range: 14.5%-22.0%). Training to prolong QE duration beyond 18% may enhance shot success.

A machine learning framework to classify musculoskeletal injury risk groups in military service members.

Background: Musculoskeletal injuries (MSKIs) are endemic in military populations. Thus, it is essential to identify and mitigate MSKI risks. Time-to-event machine learning models utilizing self-reported questionnaires or existing data (e.g., electronic health records) may aid in creating efficient risk screening tools. Methods: A total of 4,222 U.S. Army Service members completed a self-report MSKI risk screen as part of their unit's standard in-processing. Additionally, participants' MSKI and demographic data were abstracted from electronic health record data. Survival machine learning models (Cox proportional hazard regression (COX), COX with splines, conditional inference trees, and random forest) were deployed to develop a predictive model on the training data (75%; n = 2,963) for MSKI risk over varying time horizons (30, 90, 180, and 365 days) and were evaluated on the testing data (25%; n = 987). Probability of predicted risk (0.00-1.00) from the final model stratified Service members into quartiles based on MSKI risk. Results: The COX model demonstrated the best model performance over the time horizons. The time-dependent area under the curve ranged from 0.73 to 0.70 at 30 and 180 days. The index prediction accuracy (IPA) was 12% better at 180 days than the IPA of the null model (0 variables). Within the COX model, "other" race, more self-reported pain items during the movement screens, female gender, and prior MSKI demonstrated the largest hazard ratios. When predicted probability was binned into quartiles, at 180 days, the highest risk bin had an MSKI incidence rate of 2,130.82 ± 171.15 per 1,000 person-years and incidence rate ratio of 4.74 (95% confidence interval: 3.44, 6.54) compared to the lowest risk bin. Conclusion: Self-reported questionnaires and existing data can be used to create a machine learning algorithm to identify Service members' MSKI risk profiles. Further research should develop more granular Service member-specific MSKI screening tools and create MSKI risk mitigation strategies based on these screenings.

Development and Implementation of Clinical Decision Algorithms for Oncology.

Decision-making in healthcare often relies on narrative guidelines; however, these instruments are poorly accessible for supporting clinical decision-making. This study explores the application of rule-based decision logic in algorithmic modeling, emphasizing its great potential in clinical decision support and research. Integrating rule-based algorithms with existing information systems and real-world data poses a serious challenge. Integrating decision algorithms with information standards increases their effectiveness across various applications. This study outlines a method for constructing clinical decision trees (CDTs), highlighting their transparency and interpretability, using information standards as a design principle. We use the digitization of the Dutch breast cancer guideline through CDTs as a case study to exemplify their versatility and practical significance. The process step 'primary treatment' has been successfully translated from the narrative guidelines format to the anticipated ted computational format.

Designing and validating a clinical decision support algorithm for diabetic nephroprotection in older patients.

BACKGROUND: Older patients with diabetic kidney disease (DKD) often do not receive optimal pharmacological treatment. Current clinical practice guidelines (CPGs) do not incorporate the concept of personalised care. Clinical decision support (CDS) algorithms that consider both evidence and personalised care to improve patient outcomes can improve the care of older adults. The aim of this research is to design and validate a CDS algorithm for prescribing renin-angiotensin-aldosterone system inhibitors (RAASi) for older patients with diabetes. METHODS: The design of the CDS tool included the following phases: (1) gathering evidence from systematic reviews and meta-analyses of randomised clinical trials to determine the number needed to treat (NNT) and time-to-benefit (TTB) values applicable to our target population for use in the algorithm. (2) Building a list of potential cases that addressed different prescribing scenarios (starting, adding or switching to RAASi). (3) Reviewing relevant guidelines and extracting all recommendations related to prescribing RAASi for DKD. (4) Matching NNT and TTB with specific clinical cases. (5) Validating the CDS algorithm using Delphi technique. RESULTS: We created a CDS algorithm that covered 15 possible scenarios and we generated 36 personalised and nine general recommendations based on the calculated and matched NNT and TTB values and considering the patient's life expectancy and functional capacity. The algorithm was validated by experts in three rounds of Delphi study. CONCLUSION: We designed an evidence-informed CDS algorithm that integrates considerations often overlooked in CPGs. The next steps include testing the CDS algorithm in a clinical trial.

Holistic evaluation of a machine learning-based timing calibration for PET detectors under varying data sparsity.

Objective.Modern PET scanners offer precise TOF information, improving the SNR of the reconstructed images. Timing calibrations are performed to reduce the worsening effects of the system components and provide valuable TOF information. Traditional calibration procedures often provide static or linear corrections, with the drawback that higher-order skews or event-to-event corrections are not addressed. Novel research demonstrated significant improvements in the reachable timing resolutions when combining conventional calibration approaches with machine learning, with the disadvantage of extensive calibration times infeasible for a clinical application. In this work, we made the first steps towards an in-system application and analyzed the effects of varying data sparsity on a machine learning timing calibration, aiming to accelerate the calibration time. Furthermore, we demonstrated the versatility of our calibration concept by applying the procedure for the first time to analog readout technology.Approach.We modified experimentally acquired calibration data used for training regarding their statistical and spatial sparsity, mimicking reduced measurement time and variability of the training data. Trained models were tested on unseen test data, characterized by fine spatial sampling and rich statistics. In total, 80 decision tree models with the same hyperparameter settings, were trained and holistically evaluated regarding data scientific, physics-based, and PET-based quality criteria.Main results.The calibration procedure can be heavily reduced from several days to some minutes without sacrificing quality and still significantly improving the timing resolution from(304±5)psto(216±1)pscompared to conventionally used analytical calibration methods.Significance.This work serves as the first step in making the developed machine learning-based calibration suitable for an in-system application to profit from the method's capabilities on the system level. Furthermore, this work demonstrates the functionality of the methodology on detectors using analog readout technology. The proposed holistic evaluation criteria here serve as a guideline for future evaluations of machine learning-based calibration approaches.

Predicting lower limb lymphedema after cervical cancer surgery using artificial neural network and decision tree models.

PURPOSE: This study aimed to develop and validate accessible artificial neural network and decision tree models to predict the risk of lower limb lymphedema after cervical cancer surgery. METHODS: We selected 759 patients who underwent cervical cancer surgery at the Hunan Cancer Hospital from January 2010 to January 2020, collecting demographic, behavioral, clinicopathological, and disease-related data. The artificial neural network and decision tree techniques were used to construct prediction models for lower limb lymphedema after cervical cancer surgery. Then, the models' predictive efficacies were evaluated to select the optimal model using several methods, such as the area under the receiver operating characteristic curve and accuracy, sensitivity, and specificity tests. RESULTS: In the training set, the artificial neural network and decision tree model accuracies for predicting lower limb lymphedema after cervical cancer surgery were 99.80% and 88.14%, and the sensitivities 99.50% and 74.01%, respectively; the specificities were 100% and 95.20%, respectively. The area under the receiver operating characteristic curve was 1.00 for the artificial neural network and 0.92 for the decision tree model. In the test set, the artificial neural network and decision tree models' accuracies were 86.70% and 82.02%, and the sensitivities 65.70% and 67.11%, respectively; the specificities were 96.00% and 89.47%, respectively. CONCLUSION: Both models had good predictive efficacy for lower limb lymphedema after cervical cancer surgery. However, the predictive performance and stability were superior in the artificial neural network model than in the decision tree model.

Modeling groundwater redox conditions at national scale through integration of sediment color and water chemistry in a machine learning framework.

Redox conditions play a crucial role in determining the fate of many contaminants in groundwater, impacting ecosystem services vital for both the aquatic environment and human water supply. Geospatial machine learning has previously successfully modelled large-scale redox conditions. This study is the first to consolidate the complementary information provided by sediment color and water chemistry to enhance our understanding of redox conditions in Denmark. In the first step, the depth to the first redox interface is modelled using sediment color from 27,042 boreholes. In the second step, the depth of the first redox interface is compared against water chemistry data at 22,198 wells to classify redox complexity. The absence of nitrate containing water below the first redox interface is referred to as continuous redox conditions. In contrast, discontinuous redox conditions are identified by the presence of nitrate below the first redox interface. Both models are built using 20 covariate maps, encompassing diverse hydrologically relevant information. The first redox interface is modelled with a mean error of 0.0 m and a root-mean-squared error of 8.0 m. The redox complexity model attains an accuracy of 69.8 %. Results indicate a mean depth to the first redox interface of 8.6 m and a standard deviation of 6.5 m. 60 % of Denmark is classified as discontinuous, indicating complex redox conditions, predominantly collocated in clay rich glacial landscapes. Both maps, i.e., first redox interface and redox complexity are largely driven by the water table and hydrogeology. The developed maps contribute to our understanding of subsurface redox processes, supporting national-scale land-use and water management.

Enhanced multistage deep learning for diagnosing anterior disc displacement in temporomandibular joint using magnetic resonance imaging.

OBJECTIVES: This study aimed to propose a new method for the automatic diagnosis of anterior disc displacement of the temporomandibular joint (TMJ) using magnetic resonance imaging (MRI) and deep learning. By employing a multistage approach, the factors affecting the final result can be easily identified and improved. METHODS: This study introduces a multistage automatic diagnostic technique using deep learning. This process involves segmenting the target from MR images, extracting distance parameters, and classifying the diagnosis into three classes. MRI exams of 368 TMJs from 204 patients were evaluated for anterior disc displacement. In the first stage, five algorithms were used for the semantic segmentation of the disc and condyle. In the second stage, 54 distance parameters were extracted from the segments. In the third stage, a rule-based decision model was developed to link the parameters with the expert diagnosis results. RESULTS: In the first stage, DeepLabV3+ showed the best result (95% Hausdorff distance, Dice coefficient, and sensitivity of 6.47 ± 7.22, 0.84 ± 0.07, and 0.84 ± 0.09, respectively). This study used the original MRI exams as input without preprocessing and showed high segmentation performance compared with that of previous studies. In the third stage, the combination of SegNet and a random forest model yielded an accuracy of 0.89 ± 0.06. CONCLUSIONS: An algorithm was developed to automatically diagnose TMJ-anterior disc displacement using MRI. Through a multistage approach, this algorithm facilitated the improvement of results and demonstrated high accuracy from more complex inputs. Furthermore, existing radiological knowledge was applied and validated.

Impact of the restraint decision tree for physical restraint use in South Korean neurointensive care units.

BACKGROUND: Nurses in neurointensive care units (NCUs) commonly use physical restraint (PR) to prevent adverse events like unplanned removal of devices (URDs) or falls. However, PR use should be based on evidenced decisions as it has drawbacks. Unfortunately, there is a lack of research-based PR protocol to support decision-making for nurses, especially for neurocritical patients. AIM: This study developed a restraint decision tree for neurocritical patients (RDT-N) to assist nurses in making PR decisions. We assessed its effectiveness in reducing PR use and adverse events. STUDY DESIGN: This study employed a baseline and post-intervention test design at a NCU with 19 beds and 45 nurses in a tertiary hospital in a metropolitan city in South Korea. Two-hundred and thirty-seven adult patients were admitted during the study period. During the intervention, nurses were trained on the RDT-N. PR use and adverse events between the baseline and post-intervention periods were compared. RESULTS: Post-intervention, total number of restrained patients decreased (20.7%-16.3%; χ2 = 7.68, p = .006), and the average number of PR applied per restrained patient decreased (2.42-1.71; t = 5.74, p < .001). The most frequently used PR type changed from extremity cuff to mitten (χ2 = 397.62, p < .001). No falls occurred during the study periods. On the other hand, URDs at baseline were 18.67 cases per 1000 patient days in the high-risk group and 5.78 cases per 1000 patient days in the moderate-risk group; however, no URD cases were reported post-intervention. CONCLUSIONS: The RDT-N effectively reduced PR use and adverse events. Its application can enhance patient-centred care based on individual condition and potential risks in NCUs. RELEVANCE TO CLINICAL PRACTICE: Nurses can use the RDT-N to assess the need for PR in caring for neurocritical patients, reducing PR use and adverse events.

Comparative Analysis of Deterministic and Nondeterministic Decision Trees for Decision Tables from Closed Classes.

In this paper, we consider classes of decision tables with many-valued decisions closed under operations of the removal of columns, the changing of decisions, the permutation of columns, and the duplication of columns. We study relationships among three parameters of these tables: the complexity of a decision table (if we consider the depth of the decision trees, then the complexity of a decision table is the number of columns in it), the minimum complexity of a deterministic decision tree, and the minimum complexity of a nondeterministic decision tree. We consider the rough classification of functions characterizing relationships and enumerate all possible seven types of relationships.

Machine learning to mechanically assess 2D and 3D biomimetic electrospun scaffolds for tissue engineering applications: Between the predictability and the interpretability.

Currently, the use of autografts is the gold standard for the replacement of many damaged biological tissues. However, this practice presents disadvantages that can be mitigated through tissue-engineered implants. The aim of this study is to explore how machine learning can mechanically evaluate 2D and 3D polyvinyl alcohol (PVA) electrospun scaffolds (one twisted filament, 3 twisted filament and 3 twisted/braided filament scaffolds) for their use in different tissue engineering applications. Crosslinked and non-crosslinked scaffolds were fabricated and mechanically characterised, in dry/wet conditions and under longitudinal/transverse loading, using tensile testing. 28 machine learning models (ML) were used to predict the mechanical properties of the scaffolds. 4 exogenous variables (structure, environmental condition, crosslinking and direction of the load) were used to predict 2 endogenous variables (Young's modulus and ultimate tensile strength). ML models were able to identify 6 structures and testing conditions with comparable Young's modulus and ultimate tensile strength to ligamentous tissue, skin tissue, oral and nasal tissue, and renal tissue. This novel study proved that Classification and Regression Trees (CART) models were an innovative and easy to interpret tool to identify biomimetic electrospun structures; however, Cubist and Support Vector Machine (SVM) models were the most accurate, with R2 of 0.93 and 0.8, to predict the ultimate tensile strength and Young's modulus, respectively. This approach can be implemented to optimise the manufacturing process in different applications.

Systematic review of approaches to detection and classification of skin cancer using artificial intelligence: Development and prospects.

In recent years, there has been a significant improvement in the accuracy of the classification of pigmented skin lesions using artificial intelligence algorithms. Intelligent analysis and classification systems are significantly superior to visual diagnostic methods used by dermatologists and oncologists. However, the application of such systems in clinical practice is severely limited due to a lack of generalizability and risks of potential misclassification. Successful implementation of artificial intelligence-based tools into clinicopathological practice requires a comprehensive study of the effectiveness and performance of existing models, as well as further promising areas for potential research development. The purpose of this systematic review is to investigate and evaluate the accuracy of artificial intelligence technologies for detecting malignant forms of pigmented skin lesions. For the study, 10,589 scientific research and review articles were selected from electronic scientific publishers, of which 171 articles were included in the presented systematic review. All selected scientific articles are distributed according to the proposed neural network algorithms from machine learning to multimodal intelligent architectures and are described in the corresponding sections of the manuscript. This research aims to explore automated skin cancer recognition systems, from simple machine learning algorithms to multimodal ensemble systems based on advanced encoder-decoder models, visual transformers (ViT), and generative and spiking neural networks. In addition, as a result of the analysis, future directions of research, prospects, and potential for further development of automated neural network systems for classifying pigmented skin lesions are discussed.

Machine Learning-Driven Advancements in Liposomal Formulations for Targeted Drug Delivery: A Narrative Literature Review.

Nanoliposomal formulations, utilizing lipid bilayers to encapsulate therapeutic agents, hold promise for targeted drug delivery. Recent studies have explored the application of machine learning (ML) techniques in this field. This study aims to elucidate the motivations behind integrating ML into liposomal formulations, providing a nuanced understanding of its applications and highlighting potential advantages. The review begins with an overview of liposomal formulations and their role in targeted drug delivery. It then systematically progresses through current research on ML in this area, discussing the principles guiding ML adaptation for liposomal preparation and characterization. Additionally, the review proposes a conceptual model for effective ML incorporation. The review explores popular ML techniques, including ensemble learning, decision trees, instance- based learning, and neural networks. It discusses feature extraction and selection, emphasizing the influence of dataset nature and ML method choice on technique relevance. The review underscores the importance of supervised learning models for structured liposomal formulations, where labeled data is essential. It acknowledges the merits of K-fold cross-validation but notes the prevalent use of single train/test splits in liposomal formulation studies. This practice facilitates the visualization of results through 3D plots for practical interpretation. While highlighting the mean absolute error as a crucial metric, the review emphasizes consistency between predicted and actual values. It clearly demonstrates ML techniques' effectiveness in optimizing critical formulation parameters such as encapsulation efficiency, particle size, drug loading efficiency, polydispersity index, and liposomal flux. In conclusion, the review navigates the nuances of various ML algorithms, illustrating ML's role as a decision support system for liposomal formulation development. It proposes a structured framework involving experimentation, physicochemical analysis, and iterative ML model refinement through human-centered evaluation, guiding future studies. Emphasizing meticulous experimentation, interdisciplinary collaboration, and continuous validation, the review advocates seamless ML integration into liposomal drug delivery research for robust advancements. Future endeavors are encouraged to uphold these principles.

[Real-time Detection Method for Motion Artifact of Photoplethysmography Signals Based on Decision Trees].

PPG (photoplethysmography) holds significant application value in wearable and intelligent health devices. However, during the acquisition process, PPG signals can generate motion artifacts due to inevitable coupling motion, which diminishes signal quality. In response to the challenge of real-time detection of motion artifacts in PPG signals, this study analyzed the generation and significant features of PPG signal interference. Seven features were extracted from the pulse interval data, and those exhibiting notable changes were filtered using the dual-sample Kolmogorov-Smirnov test. The real-time detection of motion artifacts in PPG signals was ultimately based on decision trees. In the experimental phase, PPG signal data from 20 college students were collected to formulate the experimental dataset. The experimental results demonstrate that the proposed method achieves an average accuracy of (94.07±1.14)%, outperforming commonly used motion artifact detection algorithms in terms of accuracy and real-time performance.

Do sociodemographic risk profiles for adolescents engaging in weekly e-cigarette, cigarette, and dual product use differ?

BACKGROUND: E-cigarette use represents a contemporary mode of nicotine product use that may be changing the risk profile of participating adolescents. Understanding differences in sociodemographic characteristics of adolescents engaging in contemporary e-cigarette use and traditional cigarette use is important for effectively developing and targeting public health intervention programs. The objective of this study was to identify and compare sociodemographic risk profiles for exclusive e-cigarette use and dual-product use among a large sample of Canadian youth. METHODS: A survey of 46,666 secondary school students in the 2021-22 wave of the COMPASS study measured frequency of past month e-cigarette and cigarette use as well as age, sex, gender, racial or ethnic background, spending money, relative family affluence, and having one's own bedroom. Rates of cigarette-only, e-cigarette-only, and dual product use were calculated, and separate classification trees were run using the CART algorithm to identify sociodemographic risk profiles for weekly dual-product use and weekly e-cigarette-only use. RESULTS: Over 13% of adolescents used only e-cigarettes at least weekly, 3% engaged in weekly dual e-cigarette and cigarette use, and less than 0.5% used only cigarettes. Available spending money was a common predictor of dual-product and e-cigarette-only use. Gender diverse youth and youth with lower perceived family affluence were at higher risk for dual-product use, while white and multiethnic adolescents were at greater risk of e-cigarette-only use. Two high-risk profiles were identified for e-cigarette-only use and four high-risk profiles were identified for dual product use. CONCLUSIONS: This study used a novel modelling approach (CART) to identify combinations of sociodemographic characteristics that profile high-risk groups for exclusive e-cigarette and dual-product use. Unique risk profiles were identified, suggesting that e-cigarettes are attracting new demographics of adolescents who have not previously been considered as high-risk for traditional cigarette use.

Profile of gym-goers who do not use performance-enhancement substances.

Introduction: Currently the use of prohibited performance-enhancing substances (PES) in fitness and gym settings is a public health concern as adverse health consequences are emerging. Understanding the characteristics of gym-goers who do not use these substances could lead to an important complement to the ongoing research about risk factors for PES use. The aim of this study was to identify the profile of PES non-use in gym-goers. Methods: In total, 453 gym-goers (mean age = 35.64 years; SD = 13.08 - measure of central tendency location and measure of absolute dispersion, respectively) completed an online survey assessing sociodemographic factors, exercise characteristics, gym modalities, peers, social influence, attitudes, subjective norms, beliefs, intentions, and self-reported use of PES. Results: Decision Trees showed that being a woman, training less frequently, not practicing bodybuilding and having a negative intention to consume PES were identified as characteristics of non-users of PES. Discussion: These results may support evidence-based anti-doping interventions to prevent abusive use of PES in the fitness context.

Development and Validation of a Decision Tree Analysis Model for Predicting Home Discharge in a Convalescent Ward: A Single Institution Study.

OBJECTIVES: Accurately predicting the likelihood of inpatients' home discharge in a convalescent ward is crucial for assisting patients and families in decision-making. While logistic regression analysis has been commonly used, its complexity limits practicality in clinical settings. We focused on decision tree analysis, which is visually straightforward. This study aimed to develop and validate the accuracy of a prediction model for home discharge for inpatients in a convalescent ward using a decision tree analysis. METHODS: The cohort consisted of 651 patients admitted to our convalescent ward from 2018 to 2020. We collected data from medical records, including disease classification, sex, age, duration of acute hospitalization, discharge destination (home or nonhome), and Functional Independence Measure (FIM) subitems at admission. We divided the cohort data into training and validation sets and developed a prediction model using decision tree analysis with discharge destination as the target and other variables as predictors. The model's accuracy was validated using the validation data set. RESULTS: The decision tree model identified FIM grooming as the first single discriminator of home discharge, diverging at four points and identifying subsequent branching for the duration of acute hospitalization. The model's accuracy was 86.7%, with a sensitivity of 0.96, specificity of 0.52, positive predictive accuracy of 0.88, and negative predictive accuracy of 0.80. The area under the receiver operating characteristic curve was 0.75. CONCLUSION: The predictive model demonstrated more than moderate predictive accuracy, suggesting its utility in clinical practice. Grooming emerged as a variable with the highest explanatory power for determining home discharge.

Differentiating Pressure Ulcer Risk Levels through Interpretable Classification Models Based on Readily Measurable Indicators.

Pressure ulcers carry a significant risk in clinical practice. This paper proposes a practical and interpretable approach to estimate the risk levels of pressure ulcers using decision tree models. In order to address the common problem of imbalanced learning in nursing classification datasets, various oversampling configurations are analyzed to improve the data quality prior to modeling. The decision trees built are based on three easily identifiable and clinically relevant pressure ulcer risk indicators: mobility, activity, and skin moisture. Additionally, this research introduces a novel tabular visualization method to enhance the usability of the decision trees in clinical practice. Thus, the primary aim of this approach is to provide nursing professionals with valuable insights for assessing the potential risk levels of pressure ulcers, which could support their decision-making and allow, for example, the application of suitable preventive measures tailored to each patient's requirements. The interpretability of the models proposed and their performance, evaluated through stratified cross-validation, make them a helpful tool for nursing care in estimating the pressure ulcer risk level.

Subgroup detection in linear growth curve models with generalized linear mixed model (GLMM) trees.

Growth curve models are popular tools for studying the development of a response variable within subjects over time. Heterogeneity between subjects is common in such models, and researchers are typically interested in explaining or predicting this heterogeneity. We show how generalized linear mixed-effects model (GLMM) trees can be used to identify subgroups with different trajectories in linear growth curve models. Originally developed for clustered cross-sectional data, GLMM trees are extended here to longitudinal data. The resulting extended GLMM trees are directly applicable to growth curve models as an important special case. In simulated and real-world data, we assess performance of the extensions and compare against other partitioning methods for growth curve models. Extended GLMM trees perform more accurately than the original algorithm and LongCART, and similarly accurate compared to structural equation model (SEM) trees. In addition, GLMM trees allow for modeling both discrete and continuous time series, are less sensitive to (mis-)specification of the random-effects structure and are much faster to compute.