Model Training Research Articles

Effect of positional asymmetry palpatory models on improvement and retention of accuracy during pelvic asymmetry assessments.

Asymmetry of bony landmarks, such as the anterior superior iliac spine (ASIS) or posterior superior iliac spine (PSIS), is often utilized to identify somatic dysfunction in the pelvis. However, establishing good accuracy for these assessments can be challenging, so objective training models have been developed to enhance learning and accuracy. The objective of this study was to determine the effect of training with positional asymmetry models with objective feedback on the improvement and retention of pelvic asymmetry assessment accuracy. First-year osteopathic medical students and undergraduate interns were recruited for model training. After a basic technique demonstration, they completed a 72-question baseline assessment on the ASIS and PSIS models. Subsequent training was conducted for 5 h per week (1 h/day) for two consecutive weeks. Model accuracy was assessed four times at baseline, midpoint, final, and retention. Assessments were scored as a percent of the correctly identified asymmetries, and change scores were calculated by comparison with the previous assessment score (i.e.,baseline to midpoint, midpoint to final, final to retention) and overall (baseline to retention). Twelve students (age range, 20.3-29.2 years) participated. At baseline, overall scores were 57.6 % for ASIS and 72.9 % for PSIS models. For ASIS models, the change scores improved from baseline to midpoint (+18.9 %, p<0.001) and from midpoint to final (+6.6 %, p=0.01) but decreased from final to retention (-7.2 %, p=0.01). The overall retention scores were higher than baseline (+18.3 %, p<0.001). For PSIS models, the change scores improved from baseline to midpoint (+13.0 %, p<0.001), and the overall retention scores were higher than baseline (+15.0 %, p<0.001). Training with positional asymmetry models with objective feedback resulted in significant sustained improvements in ASIS and PSIS positional asymmetry assessment accuracy. Integration of these models into the standard medical curriculum should be considered.

Federated Learning for Ransomware-Resilient Industrial IoT: A Decentralized Framework for Secure AI at the Manufacturing Edge

Abstract Industrial Internet of Things (IIoT) deployments are facing increasing cybersecurity threats, especially with ransomware attacks on operational technology infrastructure. Traditional centralized machine learning configurations with the storage of manufacturing data in a single repository expand the attack surface area. Federated learning presents a completely new approach to conducting distributed model training across manufacturing sites with data locality. The federated learning framework uses secure aggregation protocols and encrypted communication channels to deliver intelligent systems without sending raw operational data externally. The federated model decreases the threat of ransomware propagation and exfiltration of operational data by establishing strong access control measures at the edge nodes and employing homomorphic encryption techniques. The federated approach is particularly useful in multi-site manufacturing use cases where regulatory compliance and maintaining intellectual property remain primary concerns. Demonstrations and deployments of the proposed framework in actual research problems spanning predictive maintenance, quality control, and process optimization show the model can maintain model accuracy while enhancing the operational resilience of IIoT applications. The intersection of distributed intelligence principles and cybersecurity principles provides a pathway for trustworthy AI systems in critical industrial infrastructures.

Enhancing Liver Disease Classification Using Support Vector Machine with IQR-Based Outlier Handling

Liver disease is a significant health issue that requires early and accurate diagnosis to prevent serious complications. In this study, we propose an outlier filtering approach using the Interquartile Range (IQR) to enhance the performance of the Support Vector Machine (SVM) algorithm in liver disease classification. A publicly available liver dataset consisting of 1,700 patient records with various clinical attributes was used, and the IQR method was applied to detect and remove extreme values before model training. The SVM model employed the Radial Basis Function (RBF) kernel to capture nonlinear relationships in the data. The classifier was evaluated under two conditions: without and with IQR-based outlier removal. Performance metrics including accuracy, precision, recall, F1-score, and ROC-AUC were used to assess the model. The experimental results showed that the IQR-based preprocessing improved model performance, with the accuracy increasing from 84.41% to 84.74% and the ROC-AUC score rising from 92.08% to 93.28%. Notably, the recall for the negative class improved from 84.31% to 89.76%, indicating enhanced detection of healthy patients. These findings demonstrate that outlier handling using IQR can contribute to more stable and accurate classification outcomes, especially for models that are sensitive to data irregularities such as SVM.

DualPlaqueNet with dual-branch structure and attention mechanism for carotid plaque semantic segmentation and size prediction

BackgroundWith global aging and lifestyle changes, carotid atherosclerotic plaques are a major cause of cerebrovascular disease and ischemic stroke. However, ultrasound images suffer from high noise, low contrast, and blurred edges, making it difficult for traditional image processing methods to accurately extract plaque information.ObjectiveTo establish a deep learning-based DualPlaqueNet model for semantic segmentation and size prediction of plaques in carotid ultrasound images, thereby providing comprehensive and accurate auxiliary information for clinical risk assessment and personalized diagnosis and treatment.MethodsDualPlaqueNet uses a dual-branch architecture combined with attention mechanisms and joint loss functions to optimize segmentation and regression. Notably, a multi-layer one-dimensional convolutional structure is introduced within the Efficient Channel Attention (ECA) module. The original dataset contained 287 carotid ultrasound images from patients at Zhengzhou First People’s Hospital, which were divided into training, validation, and test sets. Model training, validation, and testing were performed after preprocessing and data augmentation of the training set. Its performance was compared with three other models.ResultsIn the plaque semantic segmentation task, DualPlaqueNet outperformed the other three models across all metrics, achieving MIoU of 88.91 ± 1.027 (%), IoU (excluding background) of 88.22 ± 1.065 (%), DSC of 89.95 ± 1.102 (%), and Accuracy of 95.98 ± 0.073 (%). For plaque size prediction, this model demonstrated lower MSE and MAE, along with a higher coefficient of determination R2, proving its ability to accurately extract plaque size information from ultrasound images.ConclusionThe dual-branch design and attention mechanisms of DualPlaqueNet effectively address the challenges of ultrasound images, achieving precise segmentation and size prediction, demonstrating its potential as an auxiliary tool for future clinical applications.

Design of a machine (federated) learning based generalized model for predicting drying kinetics of foods

ABSTRACT Different from the previously published modeling techniques, the federated learning (FL) approach provides a global modeling tool for obtaining a global model for food drying processes. The main aim of this work is to design a trained federated model for modeling drying processes of different foods. Drying data for carrot, Echinacea Angustifolia, eggplant and mushroom have been used to train FL model to overcome the estimation challenges of different food drying processes using a single and food kind independent architecture. To validate the trained FL model, apple and strawberry drying data have been used. Obtained final model has been proven to ability of modeling different types of foods with higher accuracy and flexibility for future applications. Obtained FL model has proven its ability to estimate the drying characteristics of apple and strawberry that have not been used during training process with a higher accuracy R2 value of 0.9864.

High spatial resolution Raman distributed fiber sensing based on neural network dispersion compensation

Raman distributed fiber sensing (RDFS) has garnered significant research attention due to its advantages such as wide measurement range, low cost, and rapid response. However, as the sensing distance of RDFS increases, dispersion effects in sensing fiber severely degrade the spatial resolution and temperature measurement accuracy. This paper proposes a RDFS based on a 1-dimensional dilated convolutional residual neural network (1D-DCRNN). Through analysis of multiple datasets affected by dispersion and combined with theoretical models, we construct training datasets to train the 1D-DCRNN, then employ the trained model to compensate for dispersion effects to enhance system performance. Experimental results depict that at a sensing distance of 16.8 km, the temperature error is reduced from over 16.0 ℃ to below 1.0 ℃ after dispersion compensation, while the spatial resolution improves from 3.0 m to 0.4 m. This technical achieves efficient dispersion compensation without requiring physical modifications to fiber transmission lines, providing a new technical avenue for long-distance, high-precision Raman distributed fiber sensing applications.

Improving Stock Price Forecasting with Mamba Deep Learning Model

Stock price forecasting is an important challenge in financial analysis, requiring models that can efficiently process time series data and ensure computational efficiency. This study focuses on developing a deep learning model based on the Mamba architecture, using the Selective State Space model to efficiently process data, improving the accuracy of stock price forecasting, especially in volatile markets. This model also takes advantage of Mamba's features such as Kernel Fusion, Parallel Scan and Recomputation while optimizing hardware to minimize computational resources. The model training process has been applied with a unique optimization technique. Experimental results show that this model is suitable for real-time financial applications and opens up new development potential.

Methodological aspects of liver fibrosis, dystrophies, and inflammatory lesions morphological verification with artificial intelligence

Non-tumorous liver lesions, including non-alcoholic fatty liver disease, alcoholic liver disease, and drug- and toxin-induced liver injuries, represent a significant clinical and research challenge due to their high prevalence and diagnostic complexity. Traditional morphological examination methods based on visual analysis of histological slides require a high level of specialist expertise, significant time investment, and are prone to subjective factors, reducing their accuracy and reproducibility. In recent years, artificial intelligence (AI) methods, particularly deep learning technologies and convolutional neural networks, have shown significant potential for automating the diagnosis and quantitative assessment of liver morphological changes. This review analyzes publications focusing on the application of AI for morphological verification of liver pathological alterations. Special attention is given to studies employing classification and segmentation models to assess conditions such as fibrosis, hepatocyte steatosis and ballooning degeneration, inflammatory infiltration, and necrosis. The review provides a detailed overview of datasets, annotation methods, and auxiliary tools used for model training and testing. The primary data sources were histological slides, predominantly stained with hematoxylin and eosin, with additional use of histochemical and immunohistochemical staining techniques. Most studies utilized scanned digital images, which ensured a high degree of reproducibility and automation. Performance metrics varied depending on the tasks, but classification models for steatosis and segmentation models for fibrosis often achieved accuracy exceeding 90%. However, significant variability was observed in the assessment of inflammatory infiltration and necrosis. The application of AI methods helps reduce inter-operator variability, shortens diagnostic time, and shifts from subjective semi-quantitative assessment methods to objective quantitative approaches. However, the integration of these technologies into clinical practice requires further algorithm refinement, data standardization, and improved annotation methods. Thus, AI opens new opportunities for the standardization and improvement of diagnostic accuracy, which is especially critical in the context of large-scale screening and monitoring of liver disease treatment outcomes

Efficient Hotspot Detection in Solar Panels via Computer Vision and Machine Learning

Solar power generation is rapidly emerging within renewable energy due to its cost-effectiveness and ease of deployment. However, improper inspection and maintenance lead to significant damage from unnoticed solar hotspots. Even with inspections, factors like shadows, dust, and shading cause localized heat, mimicking hotspot behavior. This study emphasizes interpretability and efficiency, identifying key predictive features through feature-level and What-if Analysis. It evaluates model training and inference times to assess effectiveness in resource-limited environments, aiming to balance accuracy, generalization, and efficiency. Using Unmanned Aerial Vehicle (UAV)-acquired thermal images from five datasets, the study compares five Machine Learning (ML) models and five Deep Learning (DL) models. Explainable AI (XAI) techniques guide the analysis, with a particular focus on MPEG (Moving Picture Experts Group)-7 features for hotspot discrimination, supported by statistical validation. Medium Gaussian SVM achieved the best trade-off, with 99.3% accuracy and 18 s inference time. Feature analysis revealed blue chrominance as a strong early indicator of hotspot detection. Statistical validation across datasets confirmed the discriminative strength of MPEG-7 features. This study revisits the assumption that DL models are inherently superior, presenting an interpretable alternative for hotspot detection; highlighting the potential impact of domain mismatch. Model-level insight shows that both absolute and relative temperature variations are important in solar panel inspections. The relative decrease in “blueness” provides a crucial early indication of faults, especially in low-contrast thermal images where distinguishing normal warm areas from actual hotspot is difficult. Feature-level insight highlights how subtle changes in color composition, particularly reductions in blue components, serve as early indicators of developing anomalies.

Soursop Leaf Disease Detection With CNNs: From Training to Deployment

Soursop (Annona muricata) is a valuable tropical fruit crop that is highly susceptible to leaf diseases caused by fungal, bacterial, and viral infections. These diseases can significantly impact crop yield and quality, posing challenges for farmers, especially when early detection is delayed. This study proposes an automated solution using Convolutional Neural Networks (CNNs) to detect soursop leaf diseases through image classification. A dataset of 400 labelled leaf images, including healthy and diseased leaves (Leaf Rust, Leaf Spot, and Sooty Mold), was collected and preprocessed for the dataset. Three CNN architectures—MobileNetV2, VGG19, and ResNet50—were evaluated based on accuracy, precision, recall, and F1-score. Among them, MobileNetV2 outperformed the others, achieving 73% accuracy, 72% precision, 65% recall, and 66% F1-score and demonstrated strong consistency across classes. The best-performing model was deployed using the Flask web framework, enabling users to upload soursop leaf images and receive instant disease classification along with suggested treatments and preventive measures. This study’s novelty lies in the end-to-end pipeline, from model training to deployment via Flask, providing a ready-to-use solution for farmers.

Contrastive learning enhanced pseudo-labeling for unsupervised domain adaptation in person re-identification

Person re-identification (ReID) technology has many applications in intelligent surveillance and public safety. However, the domain difference between the source and target domains makes the generalization ability of the model extremely challenging. To reduce the dependence on labeled data, Unsupervised Domain Adaptation (UDA) methods have become an effective way to solve this problem. However, the influence of pseudo-label generated noise on model training in existing UDA methods is still significant, resulting in limited model performance on the target domain. For this reason, this paper proposes a contrast learning-based pseudo-label refinement with probabilistic uncertainty in the unsupervised domain, adapted to Person re-identification, aiming to improve the effectiveness of the unsupervised domain adapted to Person re-identification. We first enhance the feature representation of the target domain samples based on the contrast learning technique to improve their discrimination in the feature space, thereby enhancing the cross-domain migration performance of the model. Subsequently, an innovative loss function is proposed to effectively reduce the interference of label noise on the training process by refining the generation process of pseudo-labels, which solves the negative impact of inaccurate pseudo-labels on model training. Through a series of experimental validation, the method experiments on two large-scale public datasets, Market1501 and DukeMTMC, and the Rank-1 accuracy of the proposed method reaches 91.4% and 81.4%, with the mean average precision (mAP) of 79.0% and 67.9%, respectively, which proves that the research in this paper provides a good solution for the Person re-identification task with effective technical support for label noise processing and model generalization capability improvement.

Software Development for Brain Glioma Detection Using Magnetic Resonance Imaging and Deep Learning Techniques

The detection of brain gliomas is a crucial clinical challenge that requires early, accurate diagnostic methods to improve patient outcomes. This work presents the development of a deep learning-based system for glioma detection, employing an ensemble of ResNet18, VGG16, and DenseNet121 models trained with MRI images. The preprocessing involved dataset curation, image normalisation, and mask generation through K-means clustering. The trained model was integrated into a web application, allowing users to upload images and receive immediate diagnostic feedback. Experimental results demonstrate promising accuracy rates and reliable segmentation performance. This research highlights the potential of artificial intelligence (AI) to augment traditional medical imaging techniques and assist clinical diagnosis.

Generative AI enables medical image segmentation in ultra low-data regimes.

Semantic segmentation of medical images is pivotal in applications like disease diagnosis and treatment planning. While deep learning automates this task effectively, it struggles in ultra low-data regimes for the scarcity of annotated segmentation masks. To address this, we propose a generative deep learning framework that produces high-quality image-mask pairs as auxiliary training data. Unlike traditional generative models that separate data generation from model training, ours uses multi-level optimization for end-to-end data generation. This allows segmentation performance to guide the generation process, producing data tailored to improve segmentation outcomes. Our method demonstrates strong generalization across 11 medical image segmentation tasks and 19 datasets, covering various diseases, organs, and modalities. It improves performance by 10-20% (absolute) in both same- and out-of-domain settings and requires 8-20 times less training data than existing approaches. This greatly enhances the feasibility and cost-effectiveness of deep learning in data-limited medical imaging scenarios.

Deep Reinforcement Learning for Efficient Scheduling of Ground-based Astronomical Observations

Abstract Ground-based astronomical observations face inherent challenges from weather changes, target visibility window constraints, and observational requirements. Enhancing the efficiency and effectiveness of telescope operations has long been a key objective for many observatories because of the high cost of observational resources. In this study, we formalize observation scheduling as a time-dependent combinatorial optimization problem. To achieve this, we implement a pointer network with temporal attention that is capable of planning observations while accounting for time-varying factors such as moonlight interference, target altitude, and air mass, which impact the exposure time and image quality. To support the training of the deep neural network, we propose a scoring mechanism to evaluate the effectiveness of the observations, which is optimized through a refined REINFORCE algorithm with a baseline. Furthermore, an exposure time calculator and an equipment kinematic model are incorporated to dynamically estimate the time costs during the decision-making process. The simulation results demonstrated that the trained model significantly outperformed both manual scheduling and a greedy algorithm in terms of theoretical reward scores and the total number of scheduled targets. Observation experiments conducted using a dual-telescope system at Muztaga observatory further validated the superiority of our approach, demonstrating a 45.8% enhancement in the total signal-to-noise ratio across all observed targets and a 24.1% increase in the number of completed tasks under the same observing conditions.

Region Uncertainty Estimation for Medical Image Segmentation with Noisy Labels.

The success of deep learning in 3D medical image segmentation hinges on training with a large dataset of fully annotated 3D volumes, which are difficult and time-consuming to acquire. Although recent foundation models (e.g., segment anything model, SAM) can utilize sparse annotations to reduce annotation costs, segmentation tasks involving organs and tissues with blurred boundaries remain challenging. To address this issue, we propose a region uncertainty estimation framework for Computed Tomography (CT) image segmentation using noisy labels. Specifically, we propose a sample-stratified training strategy that stratifies samples according to their varying quality labels, prioritizing confident and fine-grained information at each training stage. This sample-to-voxel level processing enables more reliable supervision information to propagate to noisy label data, thus effectively mitigating the impact of noisy annotations. Moreover, we further design a boundary-guided regional uncertainty estimation module that adapts sample hierarchical training to assist in evaluating sample confidence. Experiments conducted across multiple CT datasets demonstrate the superiority of our proposed method over several competitive approaches under various noise conditions. Our proposed reliable label propagation strategy not only significantly reduces the cost of medical image annotation and robust model training but also improves the segmentation performance in scenarios with imperfect annotations, thus paving the way towards the application of medical segmentation foundation models under low-resource and remote scenarios. Code will be available at https://github.com/KHan-UJS/NoisyLabel.

Generalizable medical image enhancement using structure-preserved diffusion models

Clinical medical images often suffer from compromised quality, which negatively impacts the diagnostic process by both clinicians and AI algorithms. While GAN-based enhancement methods have been commonly developed in recent years, delicate model training is necessary due to issues with artifacts, mode collapse, and instability. Diffusion models have shown promise in generating high-quality images superior to GANs, but challenges in training data collection and domain gaps hinder applying them for medical image enhancement. Additionally, preserving fine structures in enhancing medical images with diffusion models is still an area that requires further exploration. To overcome these challenges, we propose structure-preserved diffusion models for generalizable medical image enhancement (GEDM). GEDM leverages joint supervision from enhancement and segmentation to boost structure preservation and generalizability. Specifically, synthetic data is used to collect high-low quality paired training data with structure masks, and the Laplace transform is employed to reduce domain gaps and introduce multi-scale conditions. GEDM conducts medical image enhancement and segmentation jointly, supervised by high-quality references and structure masks from the training data. Four datasets of two medical imaging modalities were collected to implement the experiments, where GEDM outperformed state-of-the-art methods in image enhancement, as well as follow-up medical analysis tasks.

Distributed Interference-Aware Power Optimization for Multi-Task Over-the-Air Federated Learning

Over-the-air federated learning (Air-FL) has emerged as a promising paradigm that integrates communication and learning, which offers significant potential to enhance model training efficiency and optimize communication resource utilization. This paper addresses the challenge of interference management in multi-cell Air-FL systems, focusing on parallel multi-task scenarios where each cell independently executes distinct training tasks. We begin by analyzing the impact of aggregation errors on local model performance within each cell, aiming to minimize the cumulative optimality gap across all cells. To this end, we formulate an optimization framework that jointly optimizes device transmit power and denoising factors. Leveraging the Pareto boundary theory, we design a centralized optimization scheme that characterizes the trade-offs in system performance. Building upon this, we propose a distributed power control optimization scheme based on interference temperature (IT). This approach decomposes the globally coupled problem into locally solvable subproblems, thereby enabling each cell to adjust its transmit power independently using only local channel state information (CSI). To tackle the non-convexity inherent in these subproblems, we first transform them into convex problems and then develop an analytical solution framework grounded in Lagrangian duality theory. Coupled with a dynamic IT update mechanism, our method iteratively approximates the Pareto optimal boundary. The simulation results demonstrate that the proposed scheme outperforms baseline methods in terms of training convergence speed, cross-cell performance balance, and test accuracy. Moreover, it achieves stable convergence within a limited number of iterations, which validates its practicality and effectiveness in multi-task edge intelligence systems.

Efficient Task Grouping Through Sample-wise Optimisation Landscape Analysis.

Shared training approaches, such as multi-task learning (MTL) and gradient-based meta-learning, are widely used in various machine learning applications, but they often suffer from negative transfer, leading to performance degradation in specific tasks. While several optimisation techniques have been developed to mitigate this issue for pre-selected task cohorts, identifying optimal task combinations for joint learning-known as task grouping-remains underexplored and computationally challenging due to the exponential growth in task combinations and the need for extensive training and evaluation cycles. This paper introduces an efficient task grouping framework designed to reduce these overwhelming computational demands of the existing methods. The proposed framework infers pairwise task similarities through a sample-wise optimisation landscape analysis, eliminating the need for the shared model training required to infer task similarities in existing methods. With task similarities acquired, a graph-based clustering algorithm is employed to pinpoint near-optimal task groups, providing an approximate yet efficient and effective solution to the originally NP-hard problem. Empirical assessments conducted on 9 different datasets highlight the effectiveness of the proposed framework, revealing a five-fold speed enhancement compared to previous state-of-the-art methods. Moreover, the framework consistently demonstrates comparable performance, confirming its remarkable efficiency and effectiveness in task grouping.

A comprehensive review of AI methods in upper extremity/limb bone fracture detection

Accurate detection of bone fractures is crucial for patient care, however, the traditional manual review of medical images like X-rays, Computed Tomography (CT) scans, Magnetic Resonance Imaging (MRIs), and ultrasounds is time-consuming and labor-intensive. The shortage of clinicians, limited access to expert radiologists, and heavy workloads increase the risk of errors, which can slow down patients recovery. Artificial Intelligence (AI) models like Faster R-CNN have shown significant diagnostic accuracy (ACC) and sensitivity (SEN), often outperforming on-call radiologists in detecting complex fracture types. For example, Faster R-CNN has achieved SEN exceeding 90% in distal radius fracture detection. However, despite these advancements, AI-driven fracture detection systems still face several challenges, including the need for extensive annotated datasets, variability in imaging quality across clinical settings, potential biases in model training, and concerns regarding the interpretability and reliability of AI-generated predictions. This review provides a comprehensive analysis of recent advancements and limitations in AI-based fracture detection, offering quantitative insights into model performance. By examining these aspects, the study highlights the importance of integrating AI systems into clinical workflows, while addressing existing barriers to their widespread adoption. This analysis underscores AI’s potential to enhance diagnostic efficiency, reduce human error, and improve patient outcomes.

Detecting schizophrenia, bipolar disorder, psychosis vulnerability and major depressive disorder from 5 minutes of online-collected speech

Psychosis poses substantial social and healthcare burdens. The analysis of speech is a promising approach for the diagnosis and monitoring of psychosis, capturing symptoms like thought disorder and flattened affect. Recent advancements in Natural Language Processing (NLP) methodologies enable the automated extraction of informative speech features, which has been leveraged for early psychosis detection and assessment of symptomology. However, critical gaps persist, including the absence of standardized sample collection protocols, small sample sizes, and a lack of multi-illness classification, limiting clinical applicability. Our study aimed to (1) identify an optimal assessment approach for the online and remote collection of speech, in the context of assessing the psychosis spectrum and evaluate whether a fully automated, speech-based machine learning (ML) pipeline can discriminate among different conditions on the schizophrenia-bipolar spectrum (SSD-BD-SPE), help-seeking comparison subjects (MDD), and healthy controls (HC) at varying layers of analysis and diagnostic complexity. We adopted online data collection methods to collect 20 min of speech and demographic information from individuals. Participants were categorized as “healthy” help-seekers (HC), having a schizophrenia-spectrum disorder (SSD), bipolar disorder (BD), major depressive disorder (MDD), or being on the psychosis spectrum with sub-clinical psychotic experiences (SPE). SPE status was determined based on self-reported clinical diagnosis and responses to the PHQ-8 and PQ-16 screening questionnaires, while other diagnoses were determined based on self-report from participants. Linguistic and paralinguistic features were extracted and ensemble learning algorithms (e.g., XGBoost) were used to train models. A 70–30% train-test split and 30-fold cross-validation was used to validate the model performance. The final analysis sample included 1140 individuals and 22,650 min of speech. Using 5 min of speech, our model could discriminate between HC and those with a serious mental illness (SSD or BD) with 86% accuracy (AUC = 0.91, Recall = 0.7, Precision = 0.98). Furthermore, our model could discern among HC, SPE, BD and SSD groups with 86% accuracy (F1 macro = 0.855, Recall Macro = 0.86, Precision Macro = 0.86). Finally, in a 5-class discrimination task including individuals with MDD, our model had 76% accuracy (F1 macro = 0.757, Recall Macro = 0.758, Precision Macro = 0.766). Our ML pipeline demonstrated disorder-specific learning, achieving excellent or good accuracy across several classification tasks. We demonstrated that the screening of mental disorders is possible via a fully automated, remote speech assessment pipeline. We tested our model on relatively high number conditions (5 classes) in the literature and in a stratified sample of psychosis spectrum, including HC, SPE, SSD and BD (4 classes). We tested our model on a large sample (N = 1150) and demonstrated best-in-class accuracy with remotely collected speech data in the psychosis spectrum, however, further clinical validation is needed to test the reliability of model performance.