Artificial Intelligence in Newborn Medicine

Early Foundations

• In the 1950s, the field of AI emerged with the seminal work of researchers such as Alan Turing, John McCarthy, Marvin Minsky, Allen Newel, and Herbert Simon.²⁴

• Alan M Turing, an English mathematician, computer scientist, cryptanalyst, philosopher, and theoretical biologist is considered the father of theoretical computer science and artificial intelligence.^25,26 During the Second World War, he was involved with the British code-breaking center that produced ultra-intelligence. His work assisted breaking the German naval ciphers, most importantly the enigma machine.

• John McCarthy coined the term “Artificial Intelligence” as early as 1955.²⁵ He organized an 8 week summer research conference in Dartmouth, New Hampshire in 1956 where 11 mathematicians and scientists laid the foundation for AI as a new scientific field. The topics of the meeting focused on computers, NLP, neural networks, theory of computation, abstraction, and creativity. The expertise of the participants was very diverse, including economists, political scientists, cognitive psychologists, computer scientists, and electrical engineers to mention a few.

• Early AI research focused on symbolic AI, which used symbolic representations and logic-based reasoning to simulate human intelligence.²⁷

• Key developments during this period included the development of expert systems, logical reasoning systems, and early forms of machine learning (ML) algorithms.²⁸

Period of Low Progress

• During the 1970s–1980s, AI research faced significant challenges and setbacks, leading to a period known as the “AI winter.”²⁹

• Financial support for AI research declined, and there was skepticism about the feasibility of achieving human-level intelligence with existing approaches.

• Despite these challenges, research continued in areas such as expert systems, NLP, and robotics.²⁴

Rise of Machine Learning (ML)

• In the 1980s and 1990s, ML emerged as a dominant paradigm in AI research.³⁰

• Researchers explored various ML techniques, including neural networks, genetic algorithms, and statistical methods, to develop AI systems that could learn from data.³¹

• Key developments during this period included the back propagation algorithm for training neural networks, the development of support vector machines (SVMs), and the rise of Bayesian methods in ML.³²

Computational Advances

• The 2000s witnessed significant advances in computational power, data availability, and algorithmic innovation, fueling rapid progress in AI research.³³

• Deep learning (DL), a subfield of ML, focused on training deep neural networks, gained prominence due to its ability to learn hierarchical representations from data.³⁴

• Breakthroughs in DL, such as the development of convolutional neural networks (CNNs) for image recognition and recurrent neural networks (RNNs) for sequence modeling, revolutionized AI applications in computer vision, NLP, and robotics.³⁵

• The IBM Watson Health was launched in 2017 after IBM-Watson defeated two human “Jeopardy” champions in 2011. The initial experience showed that AI, DL, and ML require a careful review of the quality and size of the input data, the applied algorithm, and the validity of the output results before it can be rolled out in clinical healthcare.

Interdisciplinary Collaboration and Applications (Present)

• AI research today involves interdisciplinary collaboration between computer scientists, mathematicians, engineers, neuroscientists, and domain experts from various fields.³⁶

• These technologies have been applied to a wide range of domains, including healthcare, finance, transportation, agriculture, and entertainment, and has transformed industry.³⁷

• Ongoing research in AI focuses on addressing challenges such as data privacy and governance, fairness, interpretability, and ethical considerations, as well as advancing the capabilities of AI systems to achieve or surpass human-level intelligence in complex tasks.³⁸

Overall, the development of AI has been a gradual process, driven by scientific curiosity, technological innovation, and real-world applications. It has involved contributions from researchers and practitioners across multiple disciplines. In addition, the development of AI algorithms also has significant economic implications and potential. Many for profit organizations are progressively offering commercial AI solutions.

Machine Learning (ML)

ML is focused on the development and study of statistical algorithms that can learn from data, generalize the paradigms to unseen data, and then perform tasks without explicit instructions.⁴³ ML models can be trained on large datasets of medical images, including objective imaging-based scalars such as apparent diffusion coefficients or concentration of metabolic markers; patient records; genomic data; and other healthcare data to assist in diagnosis, treatment planning, and personalized medicine.^44–50 These techniques seem to be particularly effective for tasks such as medical image analysis, disease classification, predictive modeling, and clinical decision support.⁵¹ There are three broad patterns:

• Supervised learning: These algorithms are trained on labeled datasets, where input–output pairs are used to learn the mapping between input features and target labels. Supervised learning techniques are used in medical image analysis, disease classification, and predictive modeling.⁵²

• Unsupervised learning: These involve training algorithms on unlabeled data to discover hidden patterns or structures within the data. Unsupervised learning techniques are used in clustering, anomaly detection, and dimensionality reduction tasks in medicine.⁵³

• Semi-supervised learning: These combine elements of supervised and unsupervised learning by leveraging a small amount of labeled data and a larger amount of unlabeled data to improve model performance. Semi-supervised learning techniques are used when labeled data are scarce or expensive to obtain in medical applications.⁵⁴

Deep Learning (DL)

DL is a subfield of ML. It has shown remarkable success in various medical imaging tasks, including radiology, pathology, and dermatology. DL models, such as CNNs and RNNs can learn hierarchical representations from raw data, enabling automatic identification of patterns from medical images and signals.^55,56 DL techniques are also used in medical NLP tasks, such as clinical documentation, medical transcription, and patient data extraction from electronic health records (EHRs). There are three broad patterns:

• Convolutional neural networks (CNNs): These networks are DL models designed to process structured grid-like data, such as images. CNNs are widely used in medical imaging tasks, including image classification, segmentation, object detection, and pattern recognition.⁵⁷

• Recurrent neural networks (RNNs): These are DL models with recurrent connections that enable them to capture sequential dependencies in data. RNNs are used in tasks involving sequential data, such as time-series analysis, NLP, and EHR analysis.⁵⁸

• Transformer models: These include GPT (Bidirectional Encoder Representations from Transformers) and Bidirectional Encoder Representations from Transformers (BERT), which are DL architectures designed to process data with self-attention mechanisms.⁵⁹ Transformer models are used in medical NLP tasks, including language translation, text summarization, and clinical documentation. Later in this review, we have focused on GPT and its importance in medicine and its branches such as neonatology.⁶⁰

Expert Systems

These are AI systems designed to mimic the decision-making processes of human experts in specific domains. Expert systems use rules-based approaches and knowledge representation techniques to reason and make decisions. Expert systems are used in medical diagnosis, treatment planning, and clinical decision support systems.⁶¹ These systems are particularly useful in domains where expert knowledge is well-defined and can be codified into rules or algorithms, such as pathology, radiology, and dermatology.

Natural Language Processing (NLP)

NLP involves techniques for processing and understanding human language. These algorithms can help in:

• Processing unstructured text data in medicine, including in clinical notes, research articles, and communication between healthcare providers and patients.⁶²

• Retrieval of information from medical texts. Natural language processing can then help with sentiment analysis and to summarize documents- and answer-related questions.⁶³

• Medical literature mining to develop clinical decision support systems and patient communication platforms.⁶⁴

Computer Vision

These involve techniques for analyzing and interpreting visual data, such as medical images and videos. Computer vision techniques are used in medical imaging modalities, including radiology, pathology, and dermatology, to assist in diagnosis, image interpretation, and treatment planning.⁶⁵ Pattern recognition of imaging findings has been useful in neonatal and pediatric neuroradiology.^66,67 In these programs, feeding new magnetic resonance imaging data evokes an output with a list of differential diagnoses, with probabilities and 95% confidence intervals for each entity; specific data on the MRI findings of few cases could be added to the database to improve the experience and accuracy of the program.

Reinforcement Learning (RL)

These modules involve training agents to interact with an environment and learn optimal actions through a trial-and-error process. Reinforcement learning techniques are used to plan optimized, personalized treatments adaptive therapies and medical robotics.^68,69

These are just a few examples of the types of AI that can be used in medicine. The choice of AI techniques depends on the specific task or application, as well as factors such as available data, computational resources, and regulatory considerations. Reinforcement learning techniques are still relatively new in medicine but hold promise for addressing complex and dynamic healthcare challenges. Integrating multiple AI techniques and approaches can lead to more robust and effective solutions.

Possible Advantages of Using GPT and Similar Language Models in Medical Tasks

• Natural language understanding (NLU): GPT excels in understanding and generating human-like text, making it valuable for tasks such as medical documentation, literature analysis, and patient communication. Its ability to process and generate text in natural language allows it to assist healthcare professionals in writing clinical notes, summarizing medical literature, and engaging with patients effectively.⁷¹ This may also positively impact the health and resilience of healthcare professionals by reducing administrative tasks. The GPT has also been found to be effective with minor differences in syntax and accents.

• Flexible and versatile: GPT is a highly flexible and versatile model that can be fine-tuned for a wide range of medical tasks and applications. It can be adapted to different medical specialties, languages, and healthcare settings by training it on domain-specific data and fine-tuning its parameters for specific tasks, such as medical question answering, clinical decision support, and medical image analysis.²⁸

• Large-scale pre-training: GPT is typically pre-trained on large amounts of text data from diverse sources, enabling it to learn rich representations of language and knowledge from a broad range of domains. This large-scale pre-training helps GPT capture complex linguistic patterns, domain-specific terminology, and contextual information relevant to medical tasks, making it effective in understanding and generating medical text.¹²

• Contextual understanding: GPT models, particularly those with large numbers of parameters like GPT-3, excel in capturing contextual information and understanding the nuances of language. This contextual understanding allows GPT to generate coherent and contextually relevant responses to queries, making it useful for tasks such as medical question answering, clinical documentation, and patient education.⁷²

• Continuous learning: GPT can be continuously updated and improved over time by fine-tuning it on new data and tasks. It can adapt to evolving medical knowledge, guidelines, and practices, ensuring that it remains up-to-date and relevant for medical applications.⁷³

GPT offers several advantages for certain medical tasks, but it is important to recognize that no single AI model is universally “better” than others in all contexts. Different AI models, including neural networks, ML algorithms, and statistical models, have their own strengths and limitations, and the choice of model depends on factors such as the task requirements, available data, computational resources, and ethical considerations.

Types of GPT Codes Used in Medicine

In medical AI research and applications, many versions of GPT have been tested. Three versions of GPT are best-known:

• GPT-1 was the first version of GPT released by OpenAI. It had 117 million parameters and was trained on a diverse range of internet text. GPT-1 demonstrated strong performance on various NLP tasks but was later surpassed by more sophisticated models integrating larger data sets and parameter.

• GPT-2 was a significant advancement over GPT-1, featuring 1.5 billion parameters. It was trained on a massive dataset scraped from the internet, allowing it to generate highly coherent and contextually relevant text. GPT-2 attracted attention due to concerns about its potential misuse for generating fake news or deceptive content, leading OpenAI to initially release it in a controlled manner. However, the full version was later made publicly available.

• GPT-3 is a powerful version of the model, boasting a staggering 175 billion parameters. It is the largest publicly known language model to date. GPT-3 demonstrated remarkable capabilities in generating human-like text, understanding context, and performing a wide range of NLP tasks, including translation, question answering, text completion, and more. GPT-3 has been widely used and studied by researchers and developers across various domains.⁷⁴

• GPT-4 is a multimodal large language model that was released in early 2023. It is a transformer-based model where pre-training uses both public data and “data licensed from third-party providers.” The model has been fine-tuned with reinforcement learning feedback from humans and AI for human alignment and policy compliance.

Each version of GPT has built upon the advancements of its predecessor, with improvements in model architecture, training data, and scale. The increasing size and complexity of these models have contributed to significant advancements in natural language understanding and generation, enabling a wide range of applications across industries.⁷⁵ Below is a simplified flow diagram outlining the general process of developing a GPT model for a specific application, such as medicine:

sql

Copy code

Start

|

|– Define Task and Objectives

|

|– Collect Data

| |– Gather Raw Text Data

| |– Pre-process Data (Tokenization, Cleaning, etc.)

| |– Create Training, Validation, and Test Sets

|

|– Pre-train GPT Model

| |– Initialize Model Architecture

| |– Train on Large Text Corpus (Unsupervised Learning)

| |– Fine-tune Hyperparameters

| |– Monitor Training Progress

|

|– Fine-tune GPT Model

| |– Initialize Pre-trained Model

| |– Fine-tune on Domain-specific Data (Supervised Learning)

| |– Fine-tune Task-specific Head (if applicable)

| |– Validate and Tune Hyperparameters

| |– Evaluate Performance on Validation Set

|

|– Evaluate and Test

| |– Evaluate Model Performance on Test Set

| |– Assess Metrics (Accuracy, Precision, Recall, etc.)

| |– Iterate and Refine Model (if necessary)

|

|– Deployment

| |– Integrate Model into Application or Workflow

| |– Monitor Model Performance in Real-world Settings

| |– Provide Updates and Maintenance as Needed

|

End

This flow diagram outlines the main steps involved in developing and deploying a GPT model for a specific task, such as medical text generation or medical question answering. Each step involves a series of subtasks, such as data collection, pre-processing, model training, fine-tuning, evaluation, and deployment. Additionally, there may be iterative loops where the model is refined based on performance feedback or changes in requirements.

Customized Versions of GPT

Customized versions of GPT can be developed for specific needs with appropriate expertise in ML and NLP.^76,77 Here are some general steps (Fig. 1):

• A clearly defined specific NLP task, such as text generation, text classification, sentiment analysis, language translation, or summarization.

• Collection of data: A large dataset of good quality and diversity is desirable.

• Selection of the architecture of the NLP model: A transformer-based architecture resembling GPT or other models can be used per requirements/constraints of the project.

• Pre-training of the NLP model on the dataset using self-supervised/supervised learning; this involves training the model to predict the next word(s) in a sentence given the context.

• Fine-tuning of the pre-trained model on specific examples or relevant tasks. These polishing steps improve its performance on the target task.

• Evaluation of the performance of the model on a separate validation or test dataset to assess its effectiveness and identify areas for improvement.

• Iterative refinement of the design, training, and evaluation process to improve the model. This may need change in hyperparameters or architectures, or testing with additional data.

• Once the customized language model looks satisfactory, the model can be deployed in the application or integrated into the workflow to start making predictions or generating text.

This flow diagram provides a high-level overview, and the specific details and implementation may vary depending on the task, dataset, model architecture, and research goals. Additionally, newer developments and research in the field may introduce variations or improvements to the overall process. Customized versions of GPT or a similar language model can be developed for specific needs, although it might require significant expertise in ML and NLP. Here are some general steps to consider:

Designing and training a customized language model can be a complex and resource-intensive process, requiring expertise in ML, NLP, and data engineering. Additionally, there might be a need for powerful computing resources, such as graphics processing units or Tensor Processing Units to train large-scale models effectively.

Parameters Used in the Development of GPT for Medicine

This is an evolving process.⁷⁸ Here are some general information about the parameters typically included in training GPT models for medical applications:

• Model architecture: The GPT models are based on a transformer architecture, which consists of multiple layers of self-attention mechanisms and feedforward neural networks.⁷⁹ The number of layers, hidden units, attention heads, and other architectural parameters can vary depending on the size and complexity of the model.

• Pre-training data: Models pre-trained on large amounts of text data to learn language representations can be useful.^78,80 The pre-training data may include various sources of biomedical literature, clinical notes, EHRs, drug labels, and other healthcare-related text. These datasets can range from millions to billions of tokens.

• Fine-tuning data: After pre-training, GPT models are fine-tuned on domain-specific data to adapt them to specific tasks or applications.^81,82 This fine-tuning may focus on clinical notes, medical questions/answers, EHR, and other healthcare-related documents. The size of the fine-tuning dataset can vary depending on the specific task.

• Training hyperparameters: Various hyperparameters involved in training GPT models, such as learning rate, batch size, optimization algorithm, and dropout probability are typically tuned empirically to optimize model performance on the target task or dataset.^83,84

• Task-specific head: In some implementations, GPT models may include a task-specific head or output layer tailored to the specific medical task at hand.⁸⁵ For example, in medical question answering, the output layer may consist of a softmax classifier trained to predict the most likely answer to a given medical question.⁸⁶

• Evaluation metrics: During training and evaluation, many metrics may help evaluate GPT models on medical tasks.⁸⁷ Evaluation metrics include accuracy, precision, recall, F1 score, and perplexity, depending on the specific task and evaluation criteria.

Notably, the exact parameters used in GPT development for medicine may vary with specific implementation, datasets, task requirements, and research goals. Newer developments and research in the field may introduce variations or improvements to the training and fine-tuning processes.

GPT Models Used in Medicine

GPT models have been utilized for various applications in medicine.^88,89 Although we still do not have specific GPT models for these tasks, the GPT architecture, particularly in GPT-3, has been adapted for medical use. Here are some examples:

• BioGPT: A version fine-tuned specifically for biomedical applications.⁹⁰ It has been trained on biomedical literature and clinical text to understand medical terminology, concepts, and contexts. BioGPT has been used for tasks such as medical text generation, clinical documentation, and medical question answering.

• Clinical GPT: Clinical GPT is another variant fine-tuned on EHRs, medical literature, and healthcare-related text to assist healthcare providers in clinical documentation, patient management, and decision support.⁹¹ It aims to improve clinical workflows while ensuring compliance with healthcare regulations and standards.

• GPT-M: A modified version of GPT-3 optimized for medical text generation and understanding. It has been trained on a curated dataset of medical documents, including clinical notes, research articles, and drug labels. The GPT-M can generate medical reports, summarize patient information, and answer queries accurately.

• MedGPT: A variant designed specifically for answering medical questions and knowledge retrieval.⁹² It aims to assist healthcare professionals in accessing medical knowledge efficiently to support clinical decision-making.

To summarize, GPT models have been adapted and customized for medical applications on domain-specific data and tasks. By leveraging the power of NLP, these specialized GPT models contribute to improving healthcare delivery, clinical decision support, and medical research. Many GPT and other large language models have several applications in the field of medicine, including:⁹³

• Clinical documentation and note generation: GPT can be used to assist healthcare professionals in generating clinical notes, summaries, and reports.⁹⁴ By providing relevant patient information as input, the model can generate structured, accurate documentation to improve the efficiency of medical record-keeping.

• Medical literature summarization: GPT can help summarize large volumes of medical literature, including research articles, clinical trials, and case studies.⁹⁵ This can assist healthcare professionals in staying up-to-date with the latest advancements in their field and making informed decisions about patient care.

• Medical chatbots and virtual assistants: These can provide patients with personalized medical information, answer their questions about symptoms, treatments, medications, and even schedule appointments.⁹⁶ These virtual assistants can improve patient engagement, provide continuous support all over the day, and alleviate the burden on healthcare providers.

• Clinical decision-support systems: GPT can be integrated into clinical decision support systems to assist healthcare providers in making diagnostic and treatment decisions.⁹⁷ By analyzing patient data, medical history, and symptoms, GPT can provide recommendations for appropriate diagnostic tests, treatment options, and medication dosages.

• NLP for electronic health records (EHR): GPT can help extract information from unstructured EHR. This includes identifying key medical concepts, relevant information for research or analysis, and improving the accuracy of coding and billing.⁹⁸

• Patient education and health communication: GPT can generate patient-friendly educational materials, such as articles, pamphlets, and videos, to help patients better understand their medical conditions, treatment options, and preventive care measures.⁹⁹ This can improve health literacy, patient engagement, and adherence to treatment plans.

• Drug discovery and development: GPT can help analyze biomedical literature, clinical trial data, and molecular structures to accelerate drug discovery and development processes.^100,101 By identifying drug candidates, predicting drug interactions, and optimizing drug design, GPT can contribute to advancing medical research and improving patient outcomes.¹⁰⁰

These are just a few examples of how GPT and other large language models are being utilized in the field of medicine. With continuing progress, we anticipate to see more innovative applications that leverage NLP to improve healthcare delivery and patient outcomes.^102–104

Accuracy of GPT for Medicine

The accuracy of GPT for medicine depends on several factors, including the quality of the training data, the specific medical task, and the fine-tuning process.¹⁰⁵ While GPT and similar language models have shown impressive capabilities in natural language understanding and generation, their performance in medical applications varies depending on the complexity and domain-specific nature of the tasks:

• Training data quality: The quality and quantity of the training data used to fine-tune the model are important determinants of the accuracy of GPT models.¹⁰⁶ Training data that are representative of the medical domain and cover a wide range of medical topics, specialties, and terminology are essential for achieving high accuracy.⁵²

• Task complexity: The accuracy of GPT varies depending on the complexity of the medical task.¹⁰⁷ GPT may perform well on relatively simple tasks, such as medical text generation or summarization, but may not be as effective in tasks that require deep domain knowledge or specialized expertise such as clinical management.¹⁰⁸

• Domain specificity: GPT’s accuracy for medicine is affected by its ability to understand and generate medical terminology, concepts, and contexts. GPT models fine-tuned on large volumes of medical text demonstrate better accuracy for medical tasks compared with models trained on generic text data.¹⁰⁹

• Evaluation metrics: GPT models developed for medicine are typically evaluated using standard metrics such as precision, recall, F1 score, and perplexity.¹¹⁰ However, should also be evaluated clinical relevance, interpretability, and generalizability when assessing the performance of GPT models in medical applications.

• Ethical and regulatory considerations: Besides accuracy, ethical considerations, such as patient privacy, algorithm bias, and regulatory compliance, also play a crucial role in determining the suitability of GPT for medical applications.^111,112

Overall, while GPT and similar language models have demonstrated promising capabilities in medicine, the accuracy is not guaranteed and varies depending on the specific task and application. It’s essential to carefully evaluate the performance of GPT models in medical settings before deploying these in clinical practice. Interdisciplinary collaboration between AI researchers, healthcare professionals, and domain experts is critical for leveraging GPT effectively in medicine while ensuring patient safety and quality of care.

Alternatives to GPT

There are several alternatives to GPT for NLP tasks, each with its own strengths and weaknesses.⁷⁹ Here are some notable alternatives:

• Bidirectional Encoder Representions from Transformers: BERT is another widely used transformer-based model developed by Google. Unlike GPT, which is trained in a left-to-right autoregressive manner, BERT is trained bidirectionally, allowing it to capture context from both directions. BERT is known for its effectiveness in various NLP tasks, including question answering, sentiment analysis, and named entity recognition.¹¹³

• XLNet: XLNet is a transformer-based model that extends BERT’s pre-training approach by leveraging permutation-based language modeling. XLNet can achieve state-of-the-art performance on several benchmark NLP tasks by considering all possible permutations of the input sequence during training, allowing it to capture bidirectional context more effectively.

• Transformer-XL: Transformer-XL is a variant of the above transformer model. It introduces a novel architecture that addresses the limitation of the fixed-length context window in traditional transformers by allowing for longer-term dependency modeling. Transformer-XL is particularly useful for tasks requiring long-range context understanding, such as language modeling and document summarization.

• Robustly-optimized BERT approach (RoBERTa): Robustly-optimized BERT approach is a variant of BERT developed by Facebook AI. It improves upon BERT’s pre-training objectives and training strategies to achieve better performance on downstream NLP tasks. Robustly-optimized BERT approach adopts larger batch sizes, longer training sequences, and dynamically changing masking patterns during pre-training, leading to improved robustness and generalization.

• A little BERT (ALBERT): A little BERT is a lightweight variant of BERT developed by Google Research that achieves performance that is comparable to BERT with fewer parameters. It introduces parameter-sharing techniques and factorized embedding parameterization to reduce model size and computational cost while maintaining high performance on various NLP tasks.

• T5 (Text-to-text Transfer Transformer): T5 is a transformer-based model developed by Google that frames all NLP tasks as text-to-text tasks, allowing for unified training and evaluation procedures. T5 achieves state-of-the-art performance on a wide range of NLP benchmarks by leveraging large-scale pre-training and fine-tuning on task-specific data.

These alternatives to GPT offer different approaches to modeling and training transformer-based architectures for NLP tasks. Depending on the specific requirements of a task or application, researchers and practitioners may choose one of these models based on factors such as performance, efficiency, model size, and computational resources available.

Use of GPT to Query Databases Such as PubMed

GPT can potentially be used to assist in querying large databases but there are some considerations to keep in mind:¹²¹

• Natural language interface (NLI): GPT could be employed as part of a NLI for querying databases.¹²² Users could input their queries in natural language, and GPT could help parse and interpret those queries to generate more structured search queries that are compatible with the database’s search interface.

• Query expansion and refinement: GPT could assist in refining and expanding search queries based on contextual information provided by the user.¹²³ For example, if a user provides a vague or ambiguous query, GPT could help clarify the user’s intent and suggest additional terms or concepts to include in the search.¹²⁴

• Summarization of results: After retrieving search results from PubMed or other databases, GPT could assist in summarizing and presenting the key findings or insights in a more digestible format.¹²⁵ This could include generating concise summaries of research articles, identifying relevant studies, or highlighting important keywords or concepts.

• Contextual understanding: GPT’s ability to understand context could be leveraged to improve search relevance and accuracy.¹²⁶ For example, it could utilize previous search history, user preferences, or the specific domain of interest to tailor search results more effectively.

To summarize, GPT can potentially assist in various aspects of database querying, but it is important to note that the actual querying and retrieval of data from databases like PubMed typically utilizes more specialized tools and techniques. PubMed provides its own search interface and application programming interface (API) keys for accessing its database, and many researchers often use tools like Python with libraries such as Biopython or other specialized search engines.^127–129 Overall, even though GPT can augment the querying process by providing a NLI and assisting in query refinement and result summarization, it would typically be used in conjunction with other tools and techniques for more efficient and effective database querying.

GPT to Develop Medical Manuscripts

The GPT can provide meaningful summaries with various types of images and appendices. However, the development of full manuscripts still needs work.¹³⁰

GPT to Review Medical Manuscripts

GPT can be used to assist in reviewing manuscripts, particularly in tasks such as summarization, language correction, and providing feedback on the clarity and coherence of writing.¹³¹ Here’s how GPT can be applied in manuscript review:

• Summarization: GPT can generate concise summaries of manuscripts, highlighting key findings and contributions. Reviewers can use these summaries for understanding the key points of the manuscript and provide feedback on its overall coherence and structure.¹³²

• Language correction: GPT can help identify grammatical errors, typos, and awkward phrasings in manuscripts. Reviewers can use GPT to suggest corrections and improvements to the writing style, ensuring clarity and readability.¹³²

• Feedback generation: GPT can generate feedback on various aspects of the manuscript, such as the strength of the arguments, the relevance of the literature cited, and the validity of the methodology. Reviewers can use GPT-generated feedback as a starting point for providing detailed critiques and suggestions for improvement.¹²⁶

• Plagiarism detection: While not a primary function of GPT, it can assist in identifying potential instances of plagiarism by comparing text passages in the manuscript with existing literature and databases. Reviewers can use GPT to flag suspicious similarities and recommend further investigation.¹³³

• Reviewer assistance: GPT can aid reviewers by providing additional context or background information for the manuscript. Reviewers can use GPT to look up relevant references, definitions, or explanations to enhance their understanding of the subject matter.¹²⁶

Thus, GPT can be a valuable tool in manuscript review, but it should be viewed as a supplement to human expertise and judgment rather than a replacement. Reviewers should critically evaluate the suggestions and feedback generated by GPT based on their own expertise in the field. They should also be aware of the limitations and biases of these models.

Customized versions of GPT or similar language models can be developed for specific needs, using ML, NLP, and data engineering. Additional computing resources such as GPUs or TPUs might be needed to effectively train large-scale models.

GPT to Make Images for Medical Manuscripts

GPTs are designed primarily for NLP tasks, such as text generation and understanding. However, newer generative models with embedded DL can be tailored for image generation. An example is shown in Figure 2; there are several possible versions of a photographic image (one of the two authors).^134,135

• Generative adversarial networks (GANs): These DL models are comprised of two neural networks, a generator and a discriminator, which are trained simultaneously.¹³⁶ The generator can learn to produce realistic images from random noise, while the discriminator learns to distinguish between real and fake images. Generative adversarial networks can generate high-quality images across various domains.¹³⁷ Artificial intelligence has been used to generate chatbots, which can use algorithms with AI, ML, NLU, and NLP to simulate a human conversation with text messages in a chat window. These windows can produce text, images, sounds, software, and other digital media.

• Variational autoencoders (VAEs) are another type of generative models that can learn and generate new data points by capturing the distribution of the input data. Variational autoencoders can be trained to encode input data into a lower-dimensional latent space and then decode it back into the original space. These can be used for tasks like image generation and reconstruction.¹³⁸

• Autoregressive models: GPT itself is an autoregressive model for text generation, but similar architectures can be applied to image generation tasks.⁸ Models such as PixelRNN and PixelCNN can generate images that show pixel-by-pixel congruity with previous versions.^139,140

As shown in Figures 3 to 5, current image generators can produce interesting images but there is still room for improvement. Figure 3 shows output images of the two principal authors of this article. Panel A shows an image that resembles Dr. Maheshwari but the infant seems to show congenital anomalies. In panel B, most observers did not identify the image of the physician with Dr. Huisman. Figure 4 shows the typical clinical setting of two physicians attending to a newborn baby. However, there are multiple incorrect background details and what appear to be structural abnormalities in the baby. In Figure 5, the output function failed to illustrate the requested germinal matrix hemorrhage correctly; the figure showed a whirled red color that faintly resembled an acute hemorrhage.

GPT to Tabulate Data for Medical Manuscripts

GPT tools are not specifically designed to create visualizations like bar diagrams or pie charts as these typically require manipulation of numerical data. These graphics are easier prepared using specialized software tools such as:

• Python-based libraries: Python is a general-purpose programming language; the design philosophy emphasizes code readability. It is a dynamically typed, garbage-collected program that can support multiple programming paradigms, including structured, object-oriented and functional programming.^141–143

• –Matplotlib is a widely used library that can be used for creating static, interactive, and animated visualizations. It can generate bar plots, pie charts, line plots, and scatter plots.¹⁴⁴

• –Seaborn is an advancement over Matplotlib; it provides a high-level interface for creating attractive and informative statistical graphics. These advanced visualizations can support options for further customization.¹⁴⁴

• –Plotly provides interactive plots and dashboards. It supports many chart types, including bar charts, pie charts, line charts, and scatter plots. Plotly can also be used in conjunction with Dash for building interactive web applications.^145,146

• ggplot2 is a plotting system for the R programming language inspired by the Grammar of Graphics.¹⁴⁷ It is a powerful framework for creating statistical graphics such as bar plots, pie charts, and histograms.¹⁴⁸

• Business intelligence (BI) tools like Tableau and Power BI offer interfaces for creating visualizations such as bar charts, pie charts, heatmaps, and dashboards. These tools can be customized and help connecting to data sources directly.¹⁴⁹

GPT can be useful in creating visualizations such as bar diagrams or pie charts, and generating textual descriptions and/or explanations of data visualizations.

GPT for Language Translation

GPT-based models can be used for text generation and summarization, answering of questions, and language translation.¹⁵⁰ Some transformer-based models have been specifically tailored for language translation. The GPT transformer models have been used in Google’s transformer-based translation models.¹⁵¹

Models like BERT and its variants can be fine-tuned for translation tasks. Models like T5 can be trained in a “translation mode” to perform translation tasks. Generally, GPT can generate text and understand context well, for translation tasks; specialized models can be even more effective.

GPT to Tabulate Data for Medical Manuscripts

GPT models can be tailored to assist in certain aspects related to tabulating data, even though these are not specifically designed for tabulating data or manipulating structured data formats like tables.¹⁵² These can be useful for:

• Data summarization: These can summarize textual data such as in descriptions or explanations of tabulated data to present key points or trends.

• Data interpretation: These models can help interpret textual descriptions of data as in identifying important features or insights and conveying those in natural language.

• Natural language interface: The GPT models can be used in NLI for interacting with tabular data. Users might be able to ask questions in natural language about the data, and then GPT could help parse and interpret those queries to retrieve relevant information from the dataset.

However, if the primary goal is to create or manipulate tables from raw data, other tools and techniques might be more suitable:

• Spreadsheet software: Tools like Microsoft Excel, Google Sheets, or specialized data analysis software provide powerful features for tabulating and manipulating structured data.

• Data processing libraries: Programming languages like Python offer libraries such as Pandas¹⁵³ which can help generate functions for creating, modifying, and summarizing tabular data.

• Business intelligence software like Tableau, Power BI, or Looker are designed specifically for data visualization and analysis, using interactive dashboards and tabular data sources.¹⁵⁴

Even though GPT might not be the most suitable tool for direct tabulation or statistical analysis of data to determine whether the observed differences or relationships are likely to be real or simply due to random chance.¹⁵⁵ The actual statistical analysis would need statistical software or programming languages such as R, Python with libraries like NumPy, SciPy, or statsmodels, or dedicated statistical software packages like SPSS or SAS.^156,157 Then, GPT applications can be used to query data, where an NLI could be used to parse the retrieved information, summarize/interpret the results of data analysis, identify patterns, and convey those in natural language.