Open Software Development Research Articles

Comparison of different drivers on energy systems investment dynamics to achieve the energy transition goals

The signing of the Paris Agreement represents a consensus on limiting the increase of the average global temperature compared to the preindustrial period. The European Union has set an ambitious goal of achieving climate neutrality by 2050. By 2030, the European Commission's REPowerEU plan aims to accelerate the processes of increasing the share of renewable energy sources, improving energy efficiency, reducing energy use, and, at the same time, diversifying energy sources and increasing electricity connectivity between member states.In order to achieve the stated goals, future energy systems will be based on renewable energy sources, which are characterized by production variability. Such systems require flexibilization technologies that ensure security and stability of supply. In this study, an analysis of how different fuel and technology costs, as well as policy targets, influence the pace of the energy transition dynamic was made on the case study of Croatia, as there was no such previous investigation. Therefore, the goal of the research is to identify the most prominent driver to achieve energy transition targets in Croatia, with an emphasis on the required flexibilization technologies. The investigation has been implemented in the energy planning tool H2RES, a detailed open-source long-term optimization model in a five-year step model based on scenario analysis of different RES and CO2 targets, technology costs, and fuel prices. The results reveal a need for an ambitious energy policy if the goal is to achieve full energy system decarbonization. Currently, high carbon prices and lower technology costs won't be enough to lead the energy transition, without legislative support.

Design and Development of a Flexible Manufacturing Cell Controller Using an Open-Source Communication Protocol for Interoperability

Flexible manufacturing cells provide significant advantages in low-volume mass-customization production but also induce added complexity and technical challenges in terms of integration, control, and extensibility. The variety of closed-source industrial protocols, the heterogeneous equipment, and the product’s manufacturing specifications are main points of consideration in the development of such a system. This study aims to describe the approach, from concept to implementation, for the development of the controller for a flexible manufacturing cell consisting of heterogeneous equipment in terms of functions and communication interfaces. Emphasis is put on the considerations and challenges for effective integration, extensibility, and interoperability. Scheduling and monitoring performed by the developed controller are demonstrated for a manufacturing cell producing microfluidic devices (bioMEMS) that consists of six workstations and a robot-based handling system. Communication between the system controller and the workstations was based on open-source technologies instead of proprietary software and protocols, to support interoperability and, to a considerable extent, code reusability.

Using digital technology as a platform to strengthen the continuum of care at community level for maternal, child and adolescent health in Tanzania: introducing the Afya-Tek program

Inadequate care within Tanzania’s primary health system contributes to thousands of preventable maternal and child deaths, and unwanted pregnancies each year. A key contributor is lack of coordination between three primary healthcare actors: public sector Community Health Workers (CHWs) and health facilities, and private sector Accredited Drug Dispensing Outlets (ADDOs). The Afya-Tek program aims to improve the continuity of care amongst these actors in Kibaha district, through a mobile-application based digital referral system that focuses on improving maternal, child and adolescent health. The digital system called Afya-Tek was co-created with users and beneficiaries, and utilises open-source technology in-line with existing government systems. The system guides healthcare actors with individualised decision support during client visits and recommends accurate next steps (education, treatment, or referral). From July 2020 to June 2023, a total of 241,000 individuals were enrolled in the Afya-Tek program covering 7,557 pregnant women, 6,582 postpartum women, 45,900 children, and 25,700 adolescents. CHWs have conducted a total of 626,000 home visits to provide health services, including screening clients for danger signs. This has resulted in 38,100 referrals to health facilities and 24,300 linkages to ADDOs. At the ADDO level, 48,552 clients self-presented; 33% of children with pneumonia symptoms received Amoxicillin; 34% of children with diarrhoea symptoms received ORS and zinc; and 4,203 referrals were made to nearest health facilities. Adolescents preferred services at ADDOs as a result of increased perceived privacy and confidentiality. In total, 89% of all referrals were attended by health facilities. As the first digital health program in Tanzania to demonstrate the linkage among public and private sector primary healthcare actors, Afya-Tek holds promise to improve maternal, child and adolescent health as well as for scale-up and sustainability, through incorporation of other disease conditions and integration with government’s Unified Community System (UCS).

Dual data mapping with fine-tuned large language models and asset administration shells toward interoperable knowledge representation

In the context of Industry 4.0, ensuring the compatibility of digital twins (DTs) with existing software systems in the manufacturing sector presents a significant challenge. The Asset Administration Shell (AAS), conceptualized as the standardized DT for an asset, offers a powerful framework that connects the DT with the established software infrastructure through interoperable knowledge representation. Although the IEC 63278 series specifies the AAS metamodel, it lacks a matching strategy for automating the mapping between proprietary data from existing software and AAS information models. Addressing this gap, we introduce a novel dual data mapping system (DDMS) that utilizes a fine-tuned open-source large language model (LLM) for entity matching. This system facilitates not only the mapping between existing software and AAS models but also between AAS models and standardized vocabulary dictionaries, thereby enhancing the model's semantic interoperability. A case study within the injection molding domain illustrates the practical application of DDMS for the automated creation of AAS instances, seamlessly integrating the manufacturer's existing data. Furthermore, we extensively investigate the potential of fine-tuning decode-only LLMs as generative classifiers and encoding-based classifiers for the entity matching task. To this end, we establish two AAS-specific datasets by collecting and compiling AAS-related resources. In addition, supplementary experiments are performed on general entity-matching benchmark datasets to ensure that our empirical conclusions and insights are generally applicable. The experiment results indicate that the fine-tuned generative LLM classifier achieves slightly better results, while the encoding-based classifier enables much faster inference. Furthermore, the fine-tuned LLM surpasses all state-of-the-art approaches for entity matching, including GPT-4 enhanced with in-context learning and chain of thoughts. This evidence highlights the effectiveness of the proposed DDMS in bridging the interoperability gap within DT applications, offering a scalable solution for the manufacturing industry.

Performance of ChatGPT-3.5 and ChatGPT-4 on the European Board of Urology (EBU) exams: a comparative analysis.

In the transformative era of artificial intelligence, its integration into various spheres, especially healthcare, has been promising. The objective of this study was to analyze the performance of ChatGPT, as open-source Large Language Model (LLM), in its different versions on the recent European Board of Urology (EBU) in-service assessment questions. We asked multiple choice questions of the official EBU test books to ChatGPT-3.5 and ChatGPT-4 for the following exams: exam 1 (2017-2018), exam 2 (2019-2020) and exam 3 (2021-2022). Exams were passed with ≥60% correct answers. ChatGPT-4 provided significantly more correct answers in all exams compared to the prior version 3.5 (exam 1: ChatGPT-3.5 64.3% vs. ChatGPT-4 81.6%; exam 2: 64.5% vs. 80.5%; exam 3: 56% vs. 77%, p < 0.001, respectively). Test exam 3 was the only exam ChatGPT-3.5 did not pass. Within the different subtopics, there were no significant differences of provided correct answers by ChatGPT-3.5. Concerning ChatGPT-4, the percentage in test exam 3 was significantly decreased in the subtopics Incontinence (exam 1: 81.6% vs. exam 3: 53.6%; p = 0.026) and Transplantation (exam 1: 77.8% vs. exam 3: 0%; p = 0.020). Our findings indicate that ChatGPT, especially ChatGPT-4, has the general ability to answer complex medical questions and might pass FEBU exams. Nevertheless, there is still the indispensable need for human validation of LLM answers, especially concerning health care issues.

Inductive thematic analysis of healthcare qualitative interviews using open-source large language models: How does it compare to traditional methods?

BackgroundLarge language models (LLMs) are generative artificial intelligence that have ignited much interest and discussion about their utility in clinical and research settings. Despite this interest there is sparse analysis of their use in qualitative thematic analysis comparing their current ability to that of human coding and analysis. In addition, there has been no published analysis of their use in real-world, protected health information. ObjectiveHere we fill that gap in the literature by comparing an LLM to standard human thematic analysis in real-world, semi-structured interviews of both patients and clinicians within a psychiatric setting. MethodsUsing a 70 billion parameter open-source LLM running on local hardware and advanced prompt engineering techniques, we produced themes that summarized a full corpus of interviews in minutes. Subsequently we used three different evaluation methods for quantifying similarity between themes produced by the LLM and those produced by humans. ResultsThese revealed similarities ranging from moderate to substantial (Jaccard similarity coefficients 0.44–0.69), which are promising preliminary results. ConclusionOur study demonstrates that open-source LLMs can effectively generate robust themes from qualitative data, achieving substantial similarity to human-generated themes. The validation of LLMs in thematic analysis, coupled with evaluation methodologies, highlights their potential to enhance and democratize qualitative research across diverse fields.

Stokes drift in crossing windsea and swell, and its effect on near-shore particle transport in Lofoten, Northern Norway

Transport assessments near the coast are often related to particles drifting near the surface. Such “particles” may be salmon lice, cod eggs, macro plastics or ship debris. Their drift depends on the Eulerian currents and the Stokes drift associated with the wind-generated surface wave field. The Stokes drift must be parameterized, and in doing so, one will inevitably make bulk estimates of the direction and speed of the drift for the wave spectrum. This paper implements a recently proposed parameterization of the Stokes drift vertical profile, which accounts for the effect of swell and windsea that propagate in different directions, to the open-source Lagrangian particle tracking model OpenDrift. We investigate how particles drifting in the Lofoten archipelago in Northern Norway depend on the Stokes drift and how we parameterize it. The parameterization accounting for crossing windsea and swell leads to lower residence time near the coastline than other popular parameterizations in the domain we studied.

Empirical study on fine-tuning pre-trained large language models for fault diagnosis of complex systems

With increasing complexity and interconnectivity of the modern industrial systems, effectively diagnosing faults is a core step to Prognostic and Health Management (PHM). Language models, particularly Large Language Models (LLMs) that pre-trained on massive corpora, have demonstrated remarkable capabilities in Natural Language Processing (NLP) and its related downstream tasks. However, how to leverage these models for facilitating system fault diagnosis and verify its effectiveness is rarely explored. This paper incorporates the potential comprehension ability of pre-trained LLMs, and investigates the efficacy of fine-tuning LLMs for realizing efficient system fault diagnosis. The experiments conducted in this study involve both open-source and closed-source models, and utilize a simulation and a real fault diagnosis dataset. We find that these models consistently achieve high performance across various metrics compared to the baselines. Additionally, qualitative and quantitative analysis is performed to investigate several aspects of the approach, such as the impact of dataset size, data normalization, missing values and explainability of the diagnosis, further showcasing the potential as well as limitations of the approach.

Introducing Security Mechanisms in OpenFog-Compliant Smart Buildings

Designing smart building IoT applications is a complex task. It requires efficiently integrating a broad number of heterogeneous, low-resource devices that adopt lightweight strategies. IoT frameworks, especially if they are standard-based, may help designers to scaffold the applications. OpenFog, established as IEEE 1934 standard, promotes the use of free open source (FOS) technologies and has been identified for use in smart buildings. However, smart building systems may present vulnerabilities, which can put their integrity at risk. Adopting state-of-the-art security mechanisms in this domain is critical but not trivial. It complicates the design and operation of the applications, increasing the cost of the deployed systems. In addition, difficulties may arise in finding qualified cybersecurity personnel. OpenFog identifies the security requirements of the applications, although it does not describe clearly how to implement them. This article presents a scalable architecture, based on the OpenFog reference architecture, to provide security by design in buildings of different sizes. It adopts FOS technologies over low-cost IoT devices. Moreover, it presents guidelines to help developers create secure applications, even if they are not security experts. It also proposes a selection of technologies in different layers to achieve the security dimensions defined in the X.805 ITU-T recommendation. A proof-of-concept Indoor Environment Quality (IEQ) system, based on low-cost smart nodes, was deployed in the Faculty of Engineering of Vitoria-Gasteiz to illustrate the implementation of the presented approach. The operation of the IEQ system was analyzed using software tools frequently used to find vulnerabilities in IoT applications. The use of state-of-the-art security mechanisms such as encryption, certificates, protocol selection and network partitioning/configuration in the OpenFog-based architecture improves smart building security.

Patient Flow Analysis with a Custom Simulation Engine

Patient flow simulation and analysis is one of the oldest IT -based methods used to optimize patient care processes and hospital management. During the pandemic, interest in this domain suddenly increased due to the various constraints and recommendations to reduce the likelihood of further infections in the hospital. Suddenly, metrics such as the number of patients waiting in the same area, the maximum time a patient could stay in a single room, and the minimum distance between patients became important issues to monitor and optimize. Using data and modelling concepts from various hospitals, our team developed a simulation tool that used bpmn models to define an emergency department. We then modified a single day's usual patient flow with various real-world inspired edge cases to evaluate how the simulated flow would change and which stations would become bottlenecks, where the quality of patient care would deteriorate and rooms would become overcrowded. To execute the models, we developed our own tool based on the open-source Camunda modeling tool and the Business Process Model Notation (BPMN) file format. To execute the generated models, we use our own Python-based execution environment based on the SpiffWorkflow library, which permits extensive logging and extensive customization of the attributes analysed. In addition, the modelling toolkit of Camunda was narrowed down and compiled so that it could be easily used by researchers who are not programmers. In the paper, we present both the modeling process and the scenario design process, as well as the results obtained through the runs, including the maximum waiting times during the model runs and the maximum number of patients waiting at once, which allowed us to validate the effectiveness of the framework.

VEXP: A virtual enhanced cross screen panel for off-target pharmacology alerts

We describe the development of the GSK vEXP (virtual enhanced cross screen panel) for off-target pharmacology alerts. The derivation of a panel of machine learning classification models or QSAR models (Quantitative Structure-Activity Relationship) for off-target safety assessment allows early alerting to risk factors in candidate drugs. The models are matched to an internal in-vitro biochemical screening panel described previously with some updates reported here. The extreme imbalance of some internal GSK datasets and most of the related external ChEMBL datasets is shown when considering potency thresholds relevant to in-vitro screening. The small size and bias to the active class make many ChEMBL datasets un-modellable using such thresholds. Although larger, many GSK datasets remain too imbalanced to give a performant model. The value of merging internal and external data to help rebalance datasets and improve the domain of applicability is demonstrated with improvements in model performance frequently seen on merged data. Efforts to collate public datasets with a far better balance of the missing in-actives would likely do more to improve opensource models than simply increasing dataset size. We investigate the use of moving the probability threshold and applying imbalanced learners to help overcome the imbalance problem. Both methods can produce models with improved performance when applied to imbalanced datasets. Datasets with class imbalance 95:5 % or with <100 compounds were un-modellable. Where datasets had a class imbalance of 90:10 % the imbalanced learn methods were often more performant than standard tree-based classifiers. No one classification algorithm consistently out-performed all others and our approach emphasises a standardised, automated build and evaluate approach across all classifiers to identify the best model. The application of vEXP includes ranking of hit compounds for fast prioritisation, flagging of hit series that contain systematic scaffold or functional group related risks and the confirmation that late-stage optimisation is not introducing new off-target activities in established chemical series.

Benchmarking deep Facial Expression Recognition: An extensive protocol with balanced dataset in the wild

Facial expression recognition (FER) is crucial in enhancing human-computer interaction. While current FER methods, leveraging various open-source deep learning models and training techniques, have shown promising accuracy and generalizability, their efficacy often diminishes in real-world scenarios that are not extensively studied. Addressing this gap, we introduce a novel in-the-wild balanced testing facial expression dataset designed for cross-domain validation, called BTFER. We rigorously evaluated widely utilized networks and self-designed architectures, adhering to a standardized protocol. Additionally, we explored different configurations, including input resolutions, class balance management, and pre-trained strategies, to ascertain their impact on performance. Through comprehensive testing across three major FER datasets and our in-depth cross-validation, we have ranked these network architectures and formulated a series of practical guidelines for implementing deep learning-based FER solutions in real-life applications. This paper also delves into the ethical considerations, privacy concerns, and regulatory aspects relevant to the deployment of FER technologies in sectors such as marketing, education, entertainment, and healthcare, aiming to foster responsible and effective use. The BTFER dataset and the implementation code are available in Kaggle and Github, respectively.

Automatic dendritic spine segmentation in widefield fluorescence images reveal synaptic nanostructures distribution with super-resolution imaging.

Dendritic spines are the main sites for synaptic communication in neurons, and alterations in their density, size, and shapes occur in many brain disorders. Current spine segmentation methods perform poorly in conditions with low signal-to-noise and resolution, particularly in the widefield images of thick (10 μm) brain slices. Here, we combined two open-source machine-learning models to achieve automatic 3D spine segmentation in widefield diffraction-limited fluorescence images of neurons in thick brain slices. We validated the performance by comparison with manually segmented super-resolution images of spines reconstructed from direct stochastic optical reconstruction microscopy (dSTORM). Lastly, we show an application of our approach by combining spine segmentation from diffraction-limited images with dSTORM of synaptic protein PSD-95 in the same field-of-view. This allowed us to automatically analyze and quantify the nanoscale distribution of PSD-95 inside the spine. Importantly, we found the numbers, but not the average sizes, of synaptic nanomodules and nanodomains increase with spine size.

Marker Data Enhancement For Markerless Motion Capture.

Human pose estimation models can measure movement from videos at a large scale and low cost; however, open-source pose estimation models typically detect only sparse keypoints, which leads to inaccurate joint kinematics. OpenCap, a freely available service for researchers to measure movement from videos, addresses this issue using a deep learning model-the marker enhancer-that transforms sparse keypoints into dense anatomical markers. However, OpenCap performs poorly on movements not included in the training data. Here, we create a much larger and more diverse training dataset and develop a more accurate and generalizable marker enhancer. We compiled marker-based motion capture data from 1176 subjects and synthesized 1433 hours of keypoints and anatomical markers to train the marker enhancer. We evaluated its accuracy in computing kinematics using both benchmark movement videos and synthetic data representing unseen, diverse movements. The marker enhancer improved kinematic accuracy on benchmark movements (mean error: 4.1°, max: 8.7°) compared to using video keypoints (mean: 9.6°, max: 43.1°) and OpenCap's original enhancer (mean: 5.3°, max: 11.5°). It also better generalized to unseen, diverse movements (mean: 4.1°, max: 6.7°) than OpenCap's original enhancer (mean: 40.4°, max: 252.0°). Our marker enhancer demonstrates both accuracy and generalizability across diverse movements. We integrated the marker enhancer into OpenCap, thereby offering its thousands of users more accurate measurements across a broader range of movements.

Bicycle riding environment identification for detecting traffic violation in a riding safety support information system

This paper proposes a method for identifying the bicycle riding environment using only onboard equipment. Initially, a fundamental subsystem is established for identifying the bicycle riding environment, and its functionality is validated. The findings indicate that the subsystem, utilizing an open-source trained model, can detect riding on roadways but not on sidewalks. Consequently, we emphasize the need for transfer learning, specifically using sidewalk viewpoint images, to enable the identification of bicycle riding environments. Subsequently, we conduct bicycle riding environment identification by employing a transfer learning model with manually labeled training data. The results demonstrate that after transfer learning, sidewalk riding detection, which was previously unachievable, becomes feasible. The identification rate was over 80%. Furthermore, we develop four riding environment identification algorithms, including the transfer learning model, and compare their performance across various road environments and riding conditions. Consequently, it is established that the region of interest (ROI) extension identification algorithm exhibits the highest identification performance (93% on average). As a result, this paper contributes valuable insights into the realization of bicycle riding environment identification, particularly in the context of detecting traffic violations within the riding safety support information system.

Generating open-source 3D phytoplankton models by integrating photogrammetry with scanning electron microscopy.

The community structure and ecological function of marine ecosystems are critically dependent on phytoplankton. However, our understanding of phytoplankton is limited due to the lack of detailed information on their morphology. To address this gap, we developed a framework that combines scanning electron microscopy (SEM) with photogrammetry to create realistic 3D (three-dimensional) models of phytoplankton. The workflow of this framework is demonstrated using two marine algal species, one dinoflagellate Prorocentrum micans and one diatom Halamphora sp. The resulting 3D models are made openly available and allow users to interact with phytoplankton and their complex structures virtually (digitally) and tangibly (3D printing). They also allow for surface area and biovolume calculations of phytoplankton, as well as the exploration of their light scattering properties, which are both important for ecosystem modeling. Additionally, by presenting these models to the public, it bridges the gap between scientific inquiry and education, promoting broader awareness on the importance of phytoplankton.

Synthetic datasets for open software development in rare disease research

BackgroundGlobally, researchers are working on projects aiming to enhance the availability of data for rare disease research. While data sharing remains critical, developing suitable methods is challenging due to the specific sensitivity and uniqueness of rare disease data. This creates a dilemma, as there is a lack of both methods and necessary data to create appropriate approaches initially. This work contributes to bridging this gap by providing synthetic datasets that can form the foundation for such developments.MethodsUsing a hierarchical data generation approach parameterised with publicly available statistics, we generated datasets reflecting a random sample of rare disease patients from the United States (US) population. General demographics were obtained from the US Census Bureau, while information on disease prevalence, initial diagnosis, survival rates as well as race and sex ratios were obtained from the information provided by the US Centers for Disease Control and Prevention as well as the scientific literature. The software, which we have named SynthMD, was implemented in Python as open source using libraries such as Faker for generating individual data points.ResultsWe generated three datasets focusing on three specific rare diseases with broad impact on US citizens, as well as differences in affected genders and racial groups: Sickle Cell Disease, Cystic Fibrosis, and Duchenne Muscular Dystrophy. We present the statistics used to generate the datasets and study the statistical properties of output data. The datasets, as well as the code used to generate them, are available as Open Data and Open Source Software.ConclusionThe results of our work can serve as a starting point for researchers and developers working on methods and platforms that aim to improve the availability of rare disease data. Potential applications include using the datasets for testing purposes during the implementation of information systems or tailored privacy-enhancing technologies.

A quad port dual band notch UWB MIMO antenna using hybrid decoupling structure

A four-port dual-band notch ultrawideband (UWB) multiple input and multiple output (MIMO) antenna for automotive applications is presented. Initially, a circular patch monopole antenna is designed and optimized by embedding a rectangular stub onto the radiator and lowering the ground plane to realize the UWB frequency spectrum. The dual-band frequency notch, from 4.6 to 5.9 and 6.8–8.8 GHz, is achieved using a half wavelength slot onto the feedline and a pair of H-shaped stubs across the feedline, respectively, to cover WLAN and complete x-band satellite communication. Further, the designed dual-band notch UWB antenna is transformed to a four-port MIMO antenna with interelement spacing much smaller than λ/4 (λ at 3.1 GHz) and a size of 0.46 × 0.41 × 0.01 λ3. Isolation improvement is achieved by a hybrid decoupling technique comprising a distinctive neutralization line (NL) and defected ground structure (DGS). The average isolation of the proposed design is better than 22 dB. Characteristic mode analysis of the proposed antenna is performed to analyze the specific significant mode and their contribution to the antenna radiation. The antenna design's diversity and time-domain properties are analyzed and found suitable for automotive applications. The antenna's electromagnetic interaction with the electrically large structure is studied by mounting the antenna onto an open-source automobile model, and the gain is determined to be more than 6.1 dB at the antenna's resonant frequencies.

Fluvial flood inundation and socio-economic impact model based on open data

Abstract. Fluvial floods are destructive hazards that affect millions of people worldwide each year. Forecasting flood events and their potential impacts therefore is crucial for disaster preparation and mitigation. Modeling flood inundation based on extreme value analysis of river discharges is an alternative to physical models of flood dynamics, which are computationally expensive. We present the implementation of a globally applicable, open-source fluvial flood model within a state-of-the-art risk modeling framework. It uses openly available data to rapidly compute flood inundation footprints of historic and forecasted events for the estimation of associated impacts. For the example of Pakistan, we use this flood model to compute flood depths and extents and employ it to estimate population displacement due to floods. Comparing flood extents to satellite data reveals that incorporating estimated flood protection standards does not necessarily improve the flood footprint computed by the model. We further show that, after calibrating the vulnerability of the impact model to a single event, the estimated displacement caused by past floods is in good agreement with disaster reports. Finally, we demonstrate that this calibrated model is suited for probabilistic impact-based forecasting.

Assessing the Accuracy of Artificial Intelligence Models in Scoliosis Classification and Suggested Therapeutic Approaches.

Background: Open-source artificial intelligence models (OSAIMs) are increasingly being applied in various fields, including IT and medicine, offering promising solutions for diagnostic and therapeutic interventions. In response to the growing interest in AI for clinical diagnostics, we evaluated several OSAIMs-such as ChatGPT 4, Microsoft Copilot, Gemini, PopAi, You Chat, Claude, and the specialized PMC-LLaMA 13B-assessing their abilities to classify scoliosis severity and recommend treatments based on radiological descriptions from AP radiographs. Methods: Our study employed a two-stage methodology, where descriptions of single-curve scoliosis were analyzed by AI models following their evaluation by two independent neurosurgeons. Statistical analysis involved the Shapiro-Wilk test for normality, with non-normal distributions described using medians and interquartile ranges. Inter-rater reliability was assessed using Fleiss' kappa, and performance metrics, like accuracy, sensitivity, specificity, and F1 scores, were used to evaluate the AI systems' classification accuracy. Results: The analysis indicated that although some AI systems, like ChatGPT 4, Copilot, and PopAi, accurately reflected the recommended Cobb angle ranges for disease severity and treatment, others, such as Gemini and Claude, required further calibration. Particularly, PMC-LLaMA 13B expanded the classification range for moderate scoliosis, potentially influencing clinical decisions and delaying interventions. Conclusions: These findings highlight the need for the continuous refinement of AI models to enhance their clinical applicability.