{"metadata":{"kernelspec":{"language":"python","display_name":"Python 3","name":"python3"},"language_info":{"pygments_lexer":"ipython3","nbconvert_exporter":"python","version":"3.6.4","file_extension":".py","codemirror_mode":{"name":"ipython","version":3},"name":"python","mimetype":"text/x-python"}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"markdown","source":"With the rapid advancements in artificial intelligence research and engineering, fields such as transportation, finance, education, and even the medical have witnessed significant progress. This essay, titled '**AI in Healthcare**' focuses on the advancements of artificial intelligence specifically in the healthcare field.","metadata":{}},{"cell_type":"markdown","source":"## Table of Contents\n1. [Introduction](#introduction)\n2. [Medical imaging and diagnostics](#diagnosis)\n    - [Ophthalmology](#ophthalmology)\n    - [Radiology](#radiology)\n3. [Drug Discovery and Development](#drug-discovery-and-development)\n\n4. [Ethics in AI healthcare](#ethics-in-ai-healthcare)\n5. [Conclusion](#conclusion)\n6. [References](#ref)","metadata":{}},{"cell_type":"markdown","source":"## Introduction <a name=\"introduction\"></a>\nAI, as a versatile tool, has brought about a revolutionary transformation in society across diverse sectors. Over the past years, AI has played a pivotal role in driving advancements and progress. Its influence is pervasive, extending beyond industries and permeating into our everyday lives.The impact of AI in the healthcare sector has been particularly noteworthy, showcasing its remarkable potential and significance.\n\nThis essay aims to explore several significant topics where AI has made noteworthy contributions to the healthcare field in the past years. It will delve into topics such as **Medical Imaging and Diagnosis, Drug Discovery and Development**. Additionally, the essay will address at the end the ethics AI in healthcare.\n","metadata":{}},{"cell_type":"markdown","source":"## Medical imaging and diagnostics  <a name=\"diagnosis\"></a>\n\nArtificial intelligence (AI) has revolutionised medical imaging and diagnostics in various fields, including radiology, and ophthalmology. Its impact has been significant, providing  time savings for medical staff while enhancing the accuracy, safety, and efficiency of diagnostic procedures.\n\n - **Ophthalmology** <a name=\"ophthalmology\"></a>\n\nThe diagnosis of retinal diseases, including diabetic retinopathy and glaucoma, is of paramount importance to the ophthalmological community due to the potential risk of blindness.\n\nGoogle has collaborated with a team of ophthalmologists to develop a system called \"Automated Retinal Disease Assessment.\" The objective of this system is to prevent diabetic retinopathy. It employs two versions of deep learning models: one trained on three-field color fundus photographs and the other on a single-field model [[1]](#1).\n\nHowever, there are certain limitations to this system. For instance, it does not consider important **risk factors** like blood pressure measurements[[1]](#1). Additionally, the **field of application for field analysis is limited** as the models have only been trained on color fundus photographs and did not include other imaging modalities. Consequently, this system may only detect problems within these specific fields, making it challenging to identify conditions like diabetic macular edema [[1]](#1). Lastly, there are **data limitations**. The study utilized data from screening programs where patients were referred for follow-up only when moderate or severe diabetic retinopathy was identified [[1]](#1). As a result, there might be a lack of data for detecting the disease at earlier stages.\n\nThe **EviRed** (*Evaluation intelligente de la rétinopathie diabétique*)  project [[2]](#2) explores alternative imaging modalities instead of relying solely on the conventional method of color fundus photography. Within this project, researchers have explored various techniques, including three-dimensional **(3D) structural optical coherence tomography (OCT)** [[2]](#2), an advanced eye examination that utilizes light waves, **3D OCT angiography (OCTA)** [[2]](#2), which is an OCT application for visualizing blood vessels inside the eye, and two-dimensional **(2D) line scanning ophthalmoscope (LSO)** [[2]](#2), which employs a laser to scan a narrow line across the retina, capturing multiple 2D images. The combination of these techniques has shown satisfactory results.\n\nThe researchers in this paper[[3]](#3) have been drawn to the importance of early detection and automated screening for retinal diseases, leading them to propose a framework to address these challenges. Their suggestion involves incorporating **hierarchical disease information** [[3]](#3) into the model to facilitate the learning of representations for similar diseases. Additionally, they employ **class-balanced sampling** [[3]](#3) to train the model on a balanced dataset, thereby reducing bias. Finally, they utilize **knowledge distillation** [[3]](#3) from multiple teacher models to train a unified student model. The experimental results demonstrate that their methods outperform existing techniques and enable the recognition of a wide range of retinal diseases [[3]](#3).\n\n- **Radiology**  <a name=\"radiology\"></a>\n\nThere is no doubt that cancer is one of the cruelest and deadliest diseases in the world, and one that is attracting the full attention of the medical community to achieve early diagnosis. According to the **World Health Organisation (WHO)**, **lung cancer** is one of the most deadly cancers in the world. The use of **low-dose computed tomography (LDCT)** for lung cancer screening has has been shown to be effective in reducing mortality and enabling early diagnosis [[4]](#4). However, the process of inspecting and diagnosing CT scans (a medical imaging technique used to obtain detailed internal images of the body) is time-consuming, especially with the increasing number of **CT images**. Radiologists often find themselves overwhelmed by the analysis of such vast amounts of data. As a result, the research community has made significant progress in proposing solutions to help  healthcare professionals manages their workload. This is where **computer-aided diagnosis (CAD)** systems come into play. These systems are used in medical imaging to assist in the interpretation and analysis of medical images, such as X-rays, CT scans, MRIs, and ultrasounds, ultimately reducing the workload of radiologists. CAD systems typically consist of two main modules: pulmonary nodule detection (identifying small regions of abnormal tissue observed in the lungs) and nodule malignancy evaluation. While nodule detection has shown satisfactory results, models for assessing the probability of nodule malignancy lack explanations regarding the regions detected by the models [[4]](#4).\n\n\nTo address these issues, in this this article [[4]](#4), the authors propose a system called **R2MNet**, which integrates radiological analysis and malignancy assessment. The objective is to evaluate the malignancy of nodules based on radiological analysis. The radiological analysis involves classifying nodules into different types based on the features they exhibit on CT scans. On the other hand, malignancy assessment aims to estimate the risk of a nodule developing into cancer. R2MNet comprises two sub-networks: the Radiological Analysis Network (RNet) and the Malignancy Assessment Network (MNet), each dedicated to performing their respective tasks [[4]](#4). additionally,the authors introduce the **\"Channel-Dependent Activation Mapping\" (CDAM)** technique to enhance the interpretability of the model. CDAM assists in visualizing the crucial regions in CT scans that the model focuses on when making decisions [[4]](#4).\n\n\nFurthermore,the interpretation of radiology images is documented in **radiology reports**, which contain information about the patient's health. However, these reports are unstructured and written in natural language, which is complex and ambiguous [[5]](#5). Human languages can be complex and can have different meanings that can be understood in different ways by different people. This complexity makes it challenging to use radiology reports for clinical research and other important healthcare applications. Therefore, the key idea here is to structure the information in a way that it can be effectively used.\n\nIn this article [[5]](#5), the authors developed a method to extract and structure the information in the reports, resulting in the creation of a dataset called **RadGraph**. This dataset contains entities (*such as medical terms or conditions*) and relations (*connections or associations between entities)* found in full-text chest X-ray radiology reports. The researchers provided a development dataset for testing, consisting of annotations made by board-certified radiologists to 500 radiology reports from the **MIMIC-CXR** dataset. This dataset contains a total of 14,579 entities and 10,889 relations. They also created a test dataset, which includes annotations from two independent sets of board-certified radiologists on 100 radiology reports from both the MIMIC-CXR and **CheXpert** datasets, with an equal split between them [[5]](#5).\n\nA deep learning model called **RadGraph Benchmark** was trained on the **RadGraph** dataset, and it achieved good performance. Specifically, the model achieved a micro F1 score of 0.82 on the MIMIC-CXR test set and 0.73 on the CheXpert test set for relation extraction. Micro F1 is a measure of how well the model performs in identifying relations between entities [[5]](#5).\n\nThe authors also released an inference dataset, which contains annotations automatically generated by the RadGraph Benchmark model for a large number of radiology reports. This dataset includes 220,763 reports from the MIMIC-CXR dataset (with approximately 6 million entities and 4 million relations) and 500 reports from the CheXpert dataset (with 13,783 entities and 9,908 relations). These annotations are linked to the corresponding chest radiographs [[5]](#5).\n\nHow about being able to read your own **medical reports**, such as  **X-ray results**, while you wait for a long time to see your doctor? You know, there's a new computer program called **ChatGPT** that can generate text that sounds like it was written by a human. Many people use it for work, study, and other things. In this article [[6]](#6) ,The authors wanted to see how useful ChatGPT can be in the medical field. However, since the program wasn't specifically made for medical reports, it might not always provide the right information. To understand this better, they asked **15 radiologists**, who are doctors that interpret medical images, to evaluate the quality of medical reports simplified by ChatGPT. Most of the radiologists agreed that the simplified reports were **accurate**, **complete**, and didn't pose any **potential harm** to the patients [[6]](#6). In fact, it's still important for expert doctors to review the simplified reports to make sure they are accurate and safe for the patient.\n\nTwo years ago, we lived through a nightmare with **COVID-19**, so it's important to mention it. In response to dealing with COVID-19, **screening and monitoring** patients (using chest radiographs) for the virus have become increasingly important. As a result, researchers, clinicians, and patients recognize the need to build trustworthy AI models [[7]](#7).\n\nTo classify images, AI models typically use a type of neural network called **convolutional neural networks** (CNNs). However, these models have shown some problems. They tend to be overly **cautious** or overly **confident** in making decisions, which makes them **less reliable**, especially when it comes to analyzing medical images [[7]](#7).\n\nIn this article [[7]](#7), the authors suggest using a different approach called **Vision Transformers** instead of CNNs. Vision Transformers use a mechanism called attention to learn important features from images. The idea is to examine whether Transformers can generalize well and produce reliable results compared to CNNs. The authors trained and evaluated the models both quantitatively (*using numerical measurements*) and qualitatively (*using chest radiographs*). They concluded that the approach of using attention-based feature learning with Vision Transformers shows promise in building trustworthy AI models for healthcare applications [[7]](#7).\n\n","metadata":{}},{"cell_type":"markdown","source":"## Drug Discovery and Development <a name=\"drug-discovery-and-development\"></a>\n\nThe discovery and development of a new drug, from the beginning to market release, take a long time and cost a significant amount of money [[8]](#8) [[9]](#9) [[10]](#10).  There are two main stages: drug discovery and development  [[9]](#9). \nDuring drug discovery, the aim is to find new drug structures with good properties for disease treatment. The process of discovering new drugs typically includes several steps, such as **screening potential compounds** (*evaluating a large number of molecules to identify those that have the potential to be effective as drugs*), **predicting their properties**, and **optimizing their structures**. This can involve testing their ability to bind to specific drug targets, their potency in inhibiting or activating biological processes, their stability, solubility, and potential toxicity. \n\nIn this article  [[8]](#8), the authors mention techniques and models in drug discovery for finding new drugs and predicting their properties. For instance, they discuss **high-throughput screening (HTS)**, which is a specific technique used for screening where automated systems can rapidly test thousands or even millions of molecules against specific targets or biological assays [[8]](#8).They also talk about Quantitative structure-activity relationships (QSAR), which are models used in chemistry and pharmacology to predict the properties or biological activities of a molecule based on its chemical structure [[8]](#8).Molecules are represented in different formats, such as **fixed fingerprints**, **molecular graphs**, **simplified molecular input line entry system strings \n(SMILES)**, and images [[8]](#8).Each representation format can be processed by different model architectures, including **convolutional neural networks (CNN)**, **recurrent neural networks (RNN)**, **graph neural networks (GNN)**, **variational autoencoders (VAE)**, **generative adversarial networks (GAN)**, **normalization flow models**, and **transformers** [[8]](#8). Different learning paradigms have been proposed, including **self-supervised learning** involving pre-training and fine-tuning models, **reinforcement learning** to explore the chemical space, as well as learning from few-shot examples, **metric learning**, **meta-learning**, and **active learning** [[8]](#8) .\n\nHowever, these steps are performed separately, and integrating these models can be challenging as they may be incompatible, developed using different frameworks or data, and difficult to compare [[9]](#9). This can result in inefficiencies and reduced success rates in drug discovery. In this article [[9]](#9), the authors propose a unified learning framework that combines the different stages of drug discovery into a coherent model. This framework leverages the fact that the objects of study, **compound molecules**, can naturally be represented as graphs [[9]](#9). By abstracting the problems of each stage into predictions of different molecular properties, the authors aim to unify the objectives of hit screening, property prediction, and molecule optimization. The framework combines different components such as **different loss** functions, **semi-supervised learning**, **multi-task learning**, **adversarial learning**, **generative learning**, **graph attention mechanisms**, **graph reconstruction networks**, and **matching molecular pairs** on **\"activity cliffs\"**(*This describes a situation where small structural modifications made to a molecule result in a significant variation in its biological activity*) [[9]](#9).This knowledge about \"**activity cliffs**\" , enables researchers to design new molecules with improved activity by incorporating or removing the specific structural element identified from the matching molecular pairs. \n\nFor training models aimed at discovering potential molecular candidates, **data** is often needed, which is entirely normal since data is the essence of AI. \n\nIn this article [[10]](#10), the authors indicate that the **Food and Drug Administration (FDA)** has encouraged the use of **real-world data** (RWD) in drug development. By definition, real-world data refers to data collected from sources such as **electronic medical records**, **administrative claims**, and last but not the least **billing data** that provide detailed information about patients and can be used to guide therapeutic development and patient care. The applications of the models discussed in these articles rely on the analysis of real-world data, particularly unstructured clinical notes extracted from electronic medical records. **Natural language processing techniques** driven by AI have been widely used to explore potential drug repurposing opportunities and detect drug-related adverse events [[10]](#10). Furthermore, they also enable the extraction of relevant information from clinical notes and the identification of eligible populations for clinical trials (*its the final phase of drug development before they are brought to market*).  \n\nIn the pharmaceutical industry, there are three clinical trial phases. The first two involve safety, efficacy, and efficiency evaluations of the drugs on both **healthy and sick** volunteers. The last phase aims to confirm efficacy on a large group of willing patients. Unfortunately, the field of AI for clinical trials has not reached the same level of development and documentation as the progress made in the use of artificial intelligence (AI) for drug discovery [[11]](#11). This is partly due to a lack of reference works that establish clear guidelines for issues related to clinical trials. To promote the advancement of AI in drug development, a comprehensive platform called **PyTrial** has been created [[11]](#11).  It combines different machine learning methods specifically designed for clinical trial tasks. It includes six AI tasks related to clinical trials: patient outcome prediction, trial site selection, trial outcome prediction, patient-trial matching, trial similarity search, and trial data simulation. The platform offers a unified way to load data, train models, and deploy them with just a few lines of code. It also provides over 20 ready-to-use datasets for machine learning models [[11]](#11). Additionally, **PyTrial** has a standard evaluation pipeline that measures the accuracy of prediction tasks, precision/recall for ranking tasks, as well as privacy, fidelity, and usefulness for generation tasks[[11]](#11). \n\n\nYou can find the link to **PyTrial** here: https://github.com/RyanWangZf/PyTrial.\n\n","metadata":{}},{"cell_type":"markdown","source":"## Ethics in AI healthcare <a name=\"ethics-in-ai-healthcare\"></a>\n\nAs said, AI has made significant progress and advancements in the field of healthcare. For instance, in 2022, over 300 AI/ML-based healthcare devices were approved by the FDA, and many others are currently in development, including virtual health assistants, personalized therapies, smart digital pills, and remote robotic surgeries [[12]](#12). Despite these colossal advancements, AI can be **misused** or **underutilized** when it should be the opposite [[12]](#12). Inappropriate deployment without considering the capabilities of healthcare professionals or the interests of patients can lead to misuse. AI algorithms can also unintentionally introduce biases in diagnoses due to biases present in their training datasets or produce results that are difficult for experts to understand. Therefore, it is important to address the **ethical challenges** (*the term \"ethical\" refers to the study of what is right and wrong*) related to **transparency, bias, responsibility, safety, privacy, justice, and autonomy** .\n\nLet's start by highlighting the principles of medical ethics and AI:\n\nRegarding **AI ethics**, there are two types of ethics: **normative ethics** and **virtue ethics**.\n\n- **Normative ethics**: It provides general moral rules that govern behavior. There are three main categories of ethical theories: deontology, which uses rules and processes to guide ethical decisions, consequentialism (teleology), which emphasizes outcomes and defines ethical behavior based on consequences.\n\n- **Virtue ethics**: It relies on ethical virtues (such as honesty, sympathy, prudence, sagacity, gentleness, courage) to guide actions and moral behavior.\n\nIn the medical field, there are four principles of clinical biomedical ethics: **justice, respect for autonomy, beneficence, and non-maleficence**. These principles, combined with ethical standards and practical decision-making frameworks, guide healthcare professionals.\n\nTherefore, there are many differences between the two ethics presented below:\n\n- AI does not have common goals or fiduciary duties, a history and professional standards, methods to translate principles into practice, or professional and legal accountability.\n- The development of AI in the healthcare sector is driven by the private technology sector, which prioritizes profit and cost reduction.\n- Unlike medicine, the development of AI does not have established professional standards or long-standing ethical frameworks.\n- Translating principles into practice and ensuring professional and legal accountability are also challenges in AI development.\n\nThese differences make it difficult for patients to trust AI developers to act in their best interest when developing AI solutions for smart healthcare.\n\nIn this article [[12]](#12), the authors address the ethical challenges and the need for change:\n\n- **Transparency**: AI algorithms used in healthcare are often like black boxes, making it difficult to understand how they arrive at their decisions. This lack of transparency can lead to misdiagnosis or poor outcomes for patients. It is important for developers and physicians to ensure transparency and explainability in AI systems.\n\n- **Responsibility**: When an AI system makes a decision that negatively affects a patient, it can be challenging to determine who should take responsibility. AI in healthcare involves multiple stakeholders, including hardware designers, algorithm developers, and data providers. Assigning responsibility in such complex systems becomes difficult.\n\n- **Bias**: If AI algorithms are trained on biased data, they can perpetuate healthcare disparities. For example, an algorithm trained on data primarily from white patients may not accurately diagnose illnesses in people with darker skin. Developers need to address and minimize bias in AI algorithms to ensure equitable healthcare.\n\n- **Privacy**: The use of AI in healthcare involves collecting large amounts of personal data from wearables and health apps. This raises concerns about patient privacy. It is crucial to protect patient data and ensure that AI-based healthcare products prioritize privacy.\n\n- **Safety**: AI algorithms used in healthcare must not cause harm. Incorrect diagnoses or wrong recommendations can have serious consequences. Safety regulations are necessary to ensure that AI healthcare products are effective and safe for patients.\n\n- **Autonomy**: The use of AI raises questions about informed consent. Patients have the right to know if AI was involved in their treatment decisions and what data was used to train the AI. Defining informed consent in the context of AI is important for patients to have control over their healthcare.\n\n- **Justice**: Allocating healthcare resources fairly is a challenge, and AI algorithms can complicate this further. Invisible decisions made by AI systems can impact the fair distribution of resources. It is important to ensure that AI is designed to promote justice and not reinforce existing inequalities in healthcare.\n\n\n","metadata":{}},{"cell_type":"markdown","source":"## Conclusion <a name=\"conclusion\"></a>\n\nIn summary, Artificial Intelligence is transforming healthcare by enhancing medical imaging and diagnostics, helping reduce the time and cost of drug discovery and development, and assisting medical staff in reducing their workload, thus emphasizing the need for integrating AI. \n\nHowever, it is crucial to address the ethical topics associated with AI implementation in healthcare in order to ensure responsible and equitable use.","metadata":{}},{"cell_type":"markdown","source":"**Tools used**:\n\n* 🤖 **Chat-gpt** : I only used chat-gpt to check the grammar and reformulate if necessary in the \"medical and digansotic\", \"drug discover\" and \"ethics in Ai healthcare\" sections.\n","metadata":{}},{"cell_type":"markdown","source":"<font size=\"4\">✍🏽</font> : **Abdo NASSER**\n","metadata":{}},{"cell_type":"markdown","source":"## References <a name=\"ref\"></a>\n\n<a id=\"1\">[1]</a>\nBora, A et al. (2020) Predicting the risk of developing diabetic retinopathy using deep learning . Available at: https://www.thelancet.com/journals/landig/article/PIIS2589-7500(20)30250-8/fulltext.\n\n<a id=\"2\">[2]</a>\nLi, Y. et al. (2023) Multimodal information fusion for the diagnosis of diabetic retinopathy, arXiv.org. Available at: https://arxiv.org/abs/2304.00003.  \n\n\n<a id=\"3\">[3]</a>\nJu, L. et al. (2023) Hierarchical knowledge guided learning for real-world retinal diseases recognition, arXiv.org. Available at: https://arxiv.org/abs/2111.08913.\n\n\n<a id=\"4\">[4]</a> Zheng, S. et al. (2021) Interpretative computer-aided lung cancer diagnosis: From radiology analysis to Malignancy Evaluation, arXiv.org. Available at: https://arxiv.org/abs/2102.10919.\n\n\n<a id=\"5\">[5]</a> Jain, S. et al. (2021) Radgraph: Extracting clinical entities and relations from Radiology reports, arXiv.org. Available at: https://arxiv.org/abs/2106.14463.\n\n<a id=\"6\">[6]</a> Jeblick, K. et al. (2022) Chatgpt makes medicine easy to swallow: An exploratory case study on simplified radiology reports, arXiv.org. Available at: https://arxiv.org/abs/2212.14882v1.\n\n<a id=\"7\">[7]</a> Fitzke, M. et al. (2021) RapidRead: Global deployment of state-of-the-art radiology AI for a large veterinary teleradiology practice, arXiv.org. Available at: https://arxiv.org/abs/2111.08165.\n\n<a id=\"8\">[8]</a> Deng, J. et al. (2021) Artificial Intelligence in drug discovery: Applications and techniques, arXiv.org. Available at: https://arxiv.org/abs/2106.05386.\n\n<a id=\"9\">[9]</a> Yin, Y. et al. (2023) An open unified deep graph learning framework for discovering drug leads, arXiv.org. Available at: https://arxiv.org/abs/2301.03424.\n\n\n<a id=\"10\">[10]</a> Chen, Z. et al. (2021) Applications of artificial intelligence in drug development using real-world data, arXiv.org. Available at: https://arxiv.org/abs/2101.08904.\n\n\n<a id=\"11\">[11]</a> Wang, Z. et al. (2023) Pytrial: A comprehensive platform for artificial intelligence for drug development, arXiv.org. Available at: https://arxiv.org/abs/2306.04018.\n\n<a id=\"12\">[12]</a> Pasricha, S. (2022) Ai Ethics in smart healthcare, arXiv.org. Available at: https://arxiv.org/abs/2211.06346.","metadata":{}},{"cell_type":"code","source":"import pandas as pd \n\nsubmission = pd.read_csv(\"/kaggle/input/2023-kaggle-ai-report/sample_submission.csv\")\nsubmission.head()","metadata":{"execution":{"iopub.status.busy":"2023-07-16T18:31:40.839436Z","iopub.execute_input":"2023-07-16T18:31:40.839846Z","iopub.status.idle":"2023-07-16T18:31:40.916880Z","shell.execute_reply.started":"2023-07-16T18:31:40.839815Z","shell.execute_reply":"2023-07-16T18:31:40.915651Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"submission.loc[0]['value'] = 'Other'\nsubmission.loc[1]['value'] = 'https://www.kaggle.com/code/abdonasser/ai-in-healthcare/'\nsubmission.loc[2]['value'] = 'https://www.kaggle.com/code/afamos/ai-reports'\nsubmission.loc[3]['value'] = 'https://www.kaggle.com/code/enunenun/artificial-intelligence-and-human-blindspots'\nsubmission.loc[4]['value'] = 'https://www.kaggle.com/code/amosii/artificial-intelligence-a-digital-renaissance'\nsubmission.head()","metadata":{"execution":{"iopub.status.busy":"2023-07-16T18:36:29.867604Z","iopub.execute_input":"2023-07-16T18:36:29.868069Z","iopub.status.idle":"2023-07-16T18:36:29.881772Z","shell.execute_reply.started":"2023-07-16T18:36:29.868032Z","shell.execute_reply":"2023-07-16T18:36:29.880568Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"submission.to_csv('submission.csv',index=False)","metadata":{"execution":{"iopub.status.busy":"2023-07-16T18:36:32.733919Z","iopub.execute_input":"2023-07-16T18:36:32.734347Z","iopub.status.idle":"2023-07-16T18:36:32.745816Z","shell.execute_reply.started":"2023-07-16T18:36:32.734314Z","shell.execute_reply":"2023-07-16T18:36:32.744802Z"},"trusted":true},"execution_count":null,"outputs":[]}]}