The Evolution of AI: Integrating Artificial Intelligence into Health Informatics

The views expressed are those of the author and do not necessarily reflect the views of ASPA as an organization.

By April Heyward
October 28, 2024

The antecedents of artificial intelligence in the United States can be traced back to as early as the 1930s. Researchers and practitioners in the computer science, mathematics, electrical engineering, psychology, complex systems and political science disciplines contributed to the birth of the artificial intelligence discipline. The varying disciplines yielded numerous names that would later contribute to the formal characterization of the artificial intelligence discipline. See Figure 1 for Diverging Disciplines Characterization of Pre-Artificial Intelligence. Public discourse emerged among the disciplines on “machine intelligence” during the pre-artificial intelligence era. John McCarthy (Darthmouth College) approached the Rockefeller Foundation about studying artificial intelligence in 1955. McCarthy, Marvin Minsky (Harvard University), Nathaniel Rochester (IBM Corporation) and Claude Shannon (Bell Telephone Laboratories) authored and submitted a proposal titled “A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence” to the Rockefeller Foundation in August 1955. Their primary study aim was to investigate “how to make machines use a language, form abstractions and concepts, solve kinds of problems now reserved for humans, and improve themselves.” The evolution of artificial intelligence over the last sixty-nine years has far exceeded what McCarthy, Minsky, Rochester and Shannon aimed to achieve. AI has transformed every sector of society and has created news ways of thinking, working and living. It has transformed academic disciplines that have led to spinoff academic disciplines such as computational social science, computational mathematics and computational biology. Heyward continues to underscore the evolution of artificial intelligence and will focus on the integration of large language models into the health informatics discipline.

Figure 2 – Laboratory Data Summary

Large language models (LLMs) are a type of generative AI. Generative AI is a subset of deep learning which is a subset of machine learning. LLMs can perform many tasks such as sentence completion, text and image classification, analysis, generate images and question answering. The success and effectiveness of LLMs is dependent on the training data and prompt engineering. Generative pre-trained transformer (GPT) is the most known LLM. The transformer includes an encoding component and decoding component. The user inputs a prompt (“How to create an AI assistant”) that is processed by the encoding component. The encoding component passes the prompt to the decoding component which generates the output (“Creating an AI assistant involves several steps, from defining its purpose to implementing and deploying it. Here’s a high-level overview of the process.). OpenAI released the first GPT in 2019 and there have been subsequent releases over the years. In September 2024, OpenAI released “o1 preview” which is employed for complex prompts. The user can see how long it takes for GPT to generate the output (e.g., 48 seconds) and see the steps/logic the GPT employed to generate the output.  

Figure 3 – Distribution of Hemoglobin Levels

Health informatics is concerned with the collection, storage, retrieval, communication and the optimal use of health-related data, information and knowledge. Electronic health record (EHR) systems such as Epic, Cerner and NextGen are critical to the health informatics discipline. EHR systems collect and store patient medical history, visits, laboratory tests and results, medication history, medical images, clinical notes and other information. The data and information in EHR systems leads to knowledge that contributes to diagnosis, treatment and other clinical decisions. Integrating artificial intelligence (e.g., generative AI, LLMs) into health informatics furthers knowledge and clinical decisions. Heyward analyzed a Kaggle EHR Patient Laboratory Results dataset in OpenAI’s ChatGPT-4. The EHR dataset contains laboratory results for 4,412 patients. Heyward uploaded the dataset into ChatGPT-4 and prompted ChatGPT-4 to read the dataset based on raw text prompt “Read the EHR Patient Laboratory Results”. ChatGPT-4 responded to the prompt and stated “The file contains patient laboratory results with the following columns: Hematocrit, Hemoglobin, Erythrocyte, Leukocyte, Thrombocyte, MCH, MCHC, MCV, Age, Sex and Source. Let me know if you’d like any specific analysis or insights from this data.” The first eight columns are tests relating to red blood cells and white blood cells.

Figure 4 – Comparison of Hematocrit Levels by Patient Groups 

Heyward followed up and prompted ChatGPT-4 to summarize the dataset based on the raw text prompt “Summarize the data.” ChatGPT-4 responded to the prompt and stated “I’ve provided a summary of the laboratory data, which includes statistical details like the mean, count, and frequency of categorical values. You can explore it to gain insights into the ranges and distributions of various lab results, such as hematocrit, hemoglobin and cell counts. Let me know if you’d like any further specific analysis or deeper insights from this data.” See Figure 2 for Laboratory Data Summary. An in depth analysis of the data was performed. Heyward prompted ChatGPT-4 to analyze data by age groups and sex; identify the number of patients with low, normal and high results; determine if there is correlation between tests, provide insights and to visualize test data. See Figure 3 for Distribution of Hemoglobin Levels (measures the level of protein in blood) and Figure 4 for Comparison of Hematocrit Levels (measures the volume of red blood cells) by Patient Groups as example data visualizations. AI furthers information and knowledge in Health Informatics and contributions to clinical decisions, diagnosis and treatment planning. These are examples of “how to make a machine use a language and form abstractions and concepts.”

---

Author: April Heyward is a Doctoral Candidate at Valdosta State University, ASPA SSTIG Board Member, Digital Twins for Health (DT4H) Consortium Member, and Researcher. Heyward’s research interests include, but not limited to, Artificial Intelligence, Data Science, Digital Twins for Health, Deep Learning, Generative AI, Health Informatics, Large Language Models, and Machine Learning. She can be reached at [email protected].

(No Ratings Yet)
Loading...