up:: [[DBA806 - Applied AI Innovation]] tags:: #source/course #on/AI #on/healthcare people:: [[Sreenivasa Varadharajan]] # DBA806 M5 - AI In Healthcare #### Key Discussions in Lecture Dr. Sreenivasa Varadharajan conducts a session on AI applications in healthcare, encouraging participant interaction. He covers the basics of AI and machine learning, emphasizing the challenges posed by real-world data. The transition from traditional machine learning to modern AI is explored, using the *metaphor of "Lego blocks."* The session delves into practical applications in healthcare, including [[Telemedicine]], T-surgery, and gamification for patient rehabilitation. The speaker discusses the *crucial role of AI in drug discovery, emphasizing the need for drugs to identify the correct sites of operation*. He touches on 3D mapping of protein structures in drug development. The deployment challenges of AI are highlighted, with a focus on human intelligence in model architecture. The professor mentions his experience in utilizing *3D video games as a treatment modality for amblyopia* and studying neural network modeling. The session aims to cover advances in healthcare using AI. The discussion shifts to AI applications in healthcare, specifically in the context of *diagnosing COVID-19 through X-ray images*. The professor introduces a tool called "Lob" for image processing and mentions the potential of AI surpassing human accuracy in image classification. The session also touches upon traditional approaches in machine learning, emphasizing the simplicity of predicting either continuous numbers or categories. The speaker stresses the importance of *clear and well-worded questions in AI project*s. The session concludes with an example question related to predicting the likelihood of a machine failing in the next week. The speaker discusses the challenges of *predicting machine failure* within a week and emphasizes the *importance of understanding the type of data needed to answer such questions*. They highlight various types of data, including machine details, usage history, maintenance records, and environmental conditions. The speaker mentions the *need for data engineering to convert raw data into usable variables for effective analysis*. The text introduces traditional machine learning algorithms like [[Logistic Regression]] and [[Decision Trees]] for [[Classification Problems]]. It explains how these algorithms work, using examples like age and complaints to predict machine failure. The speaker touches upon the concept of [[Ensemble Techniques]] for handling mixed data and situations where traditional methods may fail. The discussion expands to [[Unsupervised Learning]], particularly [[Clustering (ML)]] for segregating data into meaningful classes. The text mentions challenges with unstructured data, such as images, audio, and text, and introduces the idea of [[Attentional Systems (AI)]] in modern AI. The speaker briefly discusses the history of neural system research and how artificial neural systems are inspired by the human visual cortex. The text concludes by introducing the concept of [[Neural Tuning]], where *specialized cells respond to specific properties of stimuli, and complex cells combine outputs from simple cells*. The speaker highlights the hierarchical nature of the visual cortex, emphasizing the importance of understanding neural systems for modern AI applications. **Key Points:** 1. Importance of understanding data types for predicting machine failure. 2. Various data types include machine details, usage history, maintenance records, and environmental conditions. 3. Data engineering is essential to convert raw data into usable variables for analysis. 4. Introduction to traditional machine learning algorithms like logistic regression and decision trees. 5. Ensemble techniques for handling mixed data and improving classification accuracy. 6. Introduction to attentional systems in modern AI inspired by human visual cortex. 7. Historical background on neural system research and its influence on artificial neural systems. 8. Concept of neural tuning, where specialized cells respond to specific properties of stimuli. 9. Hierarchy in the visual cortex and its relevance to understanding modern AI applications. ##### Perception The speaker discusses the process by which animals or visual systems perceive objects, using the example of recognizing a car in an image. The explanation involves the *analysis of dark or light points, the combination of patterns into lines and shapes, and the identification of different features of the car*. The process is compared to the speed of computers, highlighting the impressive capabilities of the human brain despite its slower processing speed. The discussion touches on the optimization of the mammalian brain for these processes and the challenges in replicating them in computer models. The *comparison of parameters and connections between the human brain and current neural network models* is mentioned, emphasizing the complexity and scale of the brain's interconnections. The lecture briefly touches on image processing, explaining how computers perceive images as numerical values and use [[Convolutional Neural Networks (CNN)]] to process them. The transition to applying similar concepts to text and audio is introduced, discussing the representation of words as vectors based on context. Towards the end, there's a mention of [[Bayesian Probability Theory]] and probability used to predict the sequence of words. The concept of attention in neural networks is briefly explained, drawing parallels with qualitative research analysis methods. **Key Points:** - Animals or visual systems perceive objects through a complex process involving the analysis of patterns and shapes. - The human brain's processing speed is slower than computers, but its optimization allows for impressive capabilities. - Challenges exist in replicating the brain's processes in computer models, considering parameters and connections. - Image processing involves representing images as numerical values and using convolutional networks for analysis. - Attention in neural networks is compared to qualitative research analysis methods. **Attentional Mechanisms** The provided text discusses the concept of attentional mechanisms and [[Transformers (NLP)]] in the context of [[Natural Language Processing (NLP)]] and AI. The attentional mechanisms focus on specific words or phrases, allowing the model to predict the next word accurately. Transformers, built on these attentional mechanisms, convert various inputs such as images, audio, or text into vectorial representations for effective processing. The attentional mechanisms involve *query, key, and value triplets*, represented as matrices or vectors, contributing to the prediction of the next word. *Transformers use multiple layers to build embeddings that capture context information, making them powerful tools for various applications*. The text explores the idea of using transformers to convert text descriptions into audio, providing a potential solution for visual empowerment. Furthermore, the discussion touches upon the potential for abuse in AI technology and emphasizes responsible usage. It mentions the risk of [[Deepfake]] images and highlights the need for careful consideration when applying these technologies. The text also introduces the idea of *inverting the architecture to generate voice from an image*, raising concerns about potential misuse of such capabilities. **Key Points:** 1. Attentional mechanisms and transformers play a crucial role in natural language processing and AI. 2. Transformers convert various inputs, such as images, audio, or text, into vectorial representations using attentional mechanisms. 3. Query, key, and value triplets are essential components of attentional mechanisms, represented as matrices or vectors. 4. Transformers have multiple layers that build embeddings, capturing context information for accurate predictions. 5. The text explores using transformers to convert text descriptions into audio, enabling visual empowerment. 6. Responsible usage of AI technology is emphasized, considering the potential for abuse, especially in creating deepfake images. 7. Inverting the architecture can potentially generate voice from an image, raising concerns about misuse. ##### Physiotherapeutic Intervention In the traditional [[Physiotherapeutic Intervention]], the process is often expensive, painful, and inefficient due to dependencies on trained personnel, space, and resources. A major insurance company approached team for help in incorporating AI and ML to improve physiotherapy. The team gamified the process by using *augmented or virtual reality games with sensors to capture muscle activation*. An underlying AI model predicted muscle activation, providing real-time feedback to doctors. The data was stored in the cloud, allowing adaptive modification of the video game training. This *gamification improved patient compliance, reduced rehabilitation time, and enhanced efficiency*. In the healthcare industry, digitalization is exemplified by incorporating various technologies like [[Internet of Things (IoT)]], cloud, and AI/ML. This approach, illustrated in physiotherapy, shows the successful use of modern digitalization technologies for improved patient outcomes. ##### Pharmaceutical Industry and Biology In the [[Pharmaceutical Industry]], AI and predictive analytics play a crucial role. Companies like VI Nariman's are integrating AI into [[Drug Discovery]], [[Clinical Trials]], and [[Supply Chain Management]]. AI systems, such as [[AlphaFold (2020)]], have made significant strides in predicting protein folding, a critical aspect of drug design. This technology allows for the development of AI-generated drugs, like the recent one for arthritis. The application of *AI in pharmaceuticals has accelerated processes, making drug discovery more efficient*. [[Computational Modeling]] in biology, facilitated by AI, has revolutionized protein folding predictions. AlphaFold, for instance, utilizes deep learning to predict protein structures, saving computational biologists considerable effort. The ability to predict protein folding is essential for drug design, allowing for modifications to target specific areas within a protein. The discussion also touches on the significance of *3D maps in understanding protein structures* and their importance in designing drugs. The pharmaceutical industry has undergone a transformation with the integration of AI, significantly reducing the time and effort required for drug discovery and development. However, the mention of *regulatory approval indicates that despite these advancements, adherence to regulatory standards is crucial, especially in the context of clinical trials*. **Key Points:** * Traditional physiotherapeutic interventions are expensive, painful, and inefficient due to dependencies on trained personnel, space, and resources. * The gamified approach improved patient compliance, reduced rehabilitation time, and enhanced efficiency. * Digitalization in healthcare involves integrating technologies like IoT, cloud, and AI/ML for better patient outcomes. * In the pharmaceutical industry, companies are integrating AI into drug discovery, clinical trials, and supply chain management. * Computational modeling in biology, driven by AI, has revolutionized protein folding predictions, saving computational biologists significant effort. * The ability to predict protein folding is crucial for drug design, allowing modifications to target specific areas within a protein. * Despite advancements, regulatory approval remains crucial, especially in the context of clinical trials. ##### Implications **Concerns about Job Security:** - Computer scientists, especially in data science, are facing potential job challenges in the near future (2-3 years). - The emphasis is shifting towards individuals who can integrate technology with business needs. - The speaker suggests a trend where some traditional roles might become obsolete. **Role of Domain Knowledge:** - The importance of domain knowledge is highlighted, especially for data scientists and developers. - Previously, *domain-agnostic model building was crucial, but now domain knowledge is becoming essential with the rise of canned models and generative AI*. **Shift in Emphasis:** - The future emphasis for data scientists and developers is predicted to be more on deployment issues rather than model development. - The speaker suggests that the *real intelligence lies in architecting models and understanding how to combine different blocks effectively*. **Off-the-Shelf AI Models:** - Anticipation of off-the-shelf domain-specific language models and AI capabilities becoming more common in the future. - Comparison drawn to the ease of use, similar to calculators, where users can drag and drop data for quick insights. **Challenges and Future Trends:** - Discussion on the challenge of computational requirements, especially for large AI models, and the need to optimize for mobile devices. - Mention of the evolving nature of technology, with *AI applications becoming as common as calculator usage*. ##### Clinical Trials and AI Applications **Fraud Detection in Clinical Trials:** - AI systems can play a significant role in fraud detection during clinical trials. - Highlighted example of false identities and participation in multiple trials simultaneously, which AI can help prevent. **Data Fudging Detection:** - A case study involving doctors manipulating data when volunteers miss appointments, and how a simple clustering algorithm helped identify the fudged data. **Repurposing Failed Drugs:** - Success story of a company leveraging AI to analyze data from failed clinical trials and repurpose drugs for new applications. - Emphasis on the abundance of failure data compared to success data in clinical trials. **Fear of Job Irrelevance:** - Addressing the fear of job loss and becoming irrelevant in the context of AI advancements, emphasizing the ongoing importance of human expertise in certain areas. **Integration of Human Expertise with AI:** - Acknowledgment that AI applications, like chat GPT, may produce misinformation, and the role of computational scientists in validating and ensuring logical outcomes. **Anticipation of AI Progression:** - Discussion on the continuous evolution of AI capabilities and the need for human experts to adapt to higher-level roles as AI technologies advance. **Summary** The speaker discusses the impact of evolving technology on jobs, emphasizing *the need for reskilling as certain jobs become obsolete*. They highlight economic and human values as key determinants in the development of generative AI. Several deployment issues are addressed, including concerns about unemployment, relevance, and the challenge of overcoming inertia in adapting to change. **Key Points:** - Paid premium services are expected to be more refined than free ones, leading to potential job replacements due to evolving technology. - Jobs may become obsolete, requiring individuals to reskill for new opportunities. - Economic factors and human values will play a crucial role in shaping the future of generative AI. - Deployment issues involve fears of unemployment and irrelevance, leading to resistance to reskilling. - The deployment stage may face slow changes, and effective [[Knowledge Management]] is essential. - The speaker mentions the application of generative AI in the Pharma industry, focusing on *drug repurposing, design, clinical trial management, and personalized insurance premiums*. - Challenges in generative AI include the fear of becoming irrelevant, ego-related resistance to change, and inertia in adapting to new technologies. - *Blockchain technology is discussed briefly, with concerns about its potential trajectory mirroring that of nanotechnology's decline.* - Supply chain issues in healthcare include challenges in human resource scheduling and predicting requirements and replenishments. #### Summary of Slide Deck [AI in Healthcare](https://cdn.upgrad.com/uploads/production/b929d32e-2f2e-4986-9ae7-27137672ad17/AI-in-Healthcare.pptx.pdf) **Introduction:** - Background: PhD in Physics, specializing in Electrophysiology and Amblyopia (Lazy Eye). - Ambitious use of 3D video games for Amblyopia treatment. - Application of Data Science and AI in Finance, Healthcare, and Education. **Machine Learning (ML) Process:** - Traditional vs. Modern ML approaches. - Factors influencing ML success: Data quality and feature engineering. - Data Engineering: Importance of consolidated data view. **Algorithms:** - Overview of various ML algorithms (Decision Trees, Logistic Regression, K Nearest Neighbors). - Model Ensembles: Adaboost, Random Forest, Gradient Boosting Machines. **Data Science and Graphical Intuition:** - Overview of regression, optimization, clustering, and anomaly detection. - Transition from traditional ML to modern AI. **Automated Complex Cell Production:** - How computers recognize images. - Building complex cells (features) from simple cells (pixels) through weighted additions. - Introduction to self-attention mechanisms and multi-head attention. - [Illustrated: Self-Attention. A step-by-step guide to self-attention… | by Raimi Karim | Towards Data Science](https://towardsdatascience.com/illustrated-self-attention-2d627e33b20a) **Modern Day AI:** - Overview of Transformers and Universal Sentence Encoder. - Applications in text, audio, video, and image processing. - Utilizing AI for recognizing patterns in unstructured data. **Digital Technologies:** - Digitization vs. Digitalization. - Integration of [[Internet of Things (IoT)]], Cloud, AI, and [[Blockchain]] in healthcare. - Future of Physiotherapy: Cost-effective, enjoyable, real-time, and efficient. **AI in Healthcare Industry:** - Role in Pharma, Biotech, Equipment manufacturers, healthcare providers, and payers. - Application of AI in various branches of the pharmaceutical industry. **AI in Drug Discovery:** - Utilizing AI for protein folding predictions. - Encoders for molecules and their representation in vector form. - Quantum computing for brute force searching and drug discovery. **Clinical Trials:** - Phases of clinical trials and their objectives. - Clinical trial fraud detection: Importance and use cases. - Innovative uses of clinical trial data in building successful businesses and mentoring systems. **Deployment Issues and People:** - Challenges in deploying AI in hospitals. - Addressing issues like planning, diagnostics, and overcoming human factors like fear, ego, and inertia. **Diagnostics in ML Era:** - Supporting doctors through AI in diagnostics. - Trade-off between accuracy and explainability. - Framework for healthcare applications and potential considerations. **Decision Support or Automation:** - Differentiating between decision support and automation. - Considering factors like clinical objectives, time-saving, and life-saving potential. - Highlighting the role of AI in medication adherence.