Medicine is a continuously evolving, dynamic field aimed at improving patient care. Hospitals, hospital management, doctors, nurses, other healthcare workers, insurance companies, pathology laboratories, imaging, pharmacy, pharmaceutical industries, research, and many other elements are present in a smoothly functioning healthcare ecosystem. Artificial Intelligence (AI) has the potential to improve accuracy, precision, and outcomes in multiple aspects of this ecosystem. It also has the ability to reduce the burden on the healthcare system by assisting in laboratory diagnosis, clinical diagnosis, imaging analysis, research studies, financial management, documentation, streamlining workflow, and much more. Machine Learning (ML), Deep Learning (DL), and Natural Language Processing (NLP) are different AI approaches applied in the healthcare sector.

The application of different AI techniques in the healthcare sector is primarily determined by the type of data to be interpreted. Healthcare providers, insurance companies, pharmaceutical firms, and research organizations are some of the important sources of healthcare data. Data is primarily of two types: structured and unstructured. Structured data is consistent and well organized (e.g., blood glucose values of patients taking part in a research study). Unstructured data is inconsistent and may vary widely amongst itself (e.g., human language, imaging, signals such as ECG). After plotting data on a correct timeline, eliminating biases, and converting it into a format interpretable by the corresponding AI application, the data is ready to be used for training the respective AI model.

Different kinds of data require different AI architectures:

1. Imaging – Common modes of grayscale (black and white) imaging in medicine include X-ray, Ultrasonography (USG), Computed Tomography (CT), and Magnetic Resonance Imaging (MRI). Endoscopy and laparoscopy are the most common modes of color imaging. Neural networks are the method of choice for interpreting this type of data. Convolutional Neural Network (CNN) is a type of neural network designed for interpreting images. The CNNs have many architectures (ResNet, VGG, GoogLeNet, UNet) that can be used and modified according to the type of problem at hand. An important concept here is the application of overlapping numerical values to the intensity of the image in a grid-like pattern to represent the relative intensity of a color. Grayscale images consist of only one grid of numerical values. Color images consist of three grids layered on one another representing Red, Green, and Blue color intensities. Image data can be used for predicting the morbidity and mortality of patients, classifying and triaging patients, and improving diagnostic accuracy.

2. Human Language – This is critical clinical data, primarily obtained from the notes of physicians, pharmacy, and nursing staff. Human language is highly complex to understand due to the changing context of different words, the relative language proficiency of the person writing the notes, the vast vocabulary and rules of grammar in any language, and the subtle differences in the meaning of seemingly similar words. The application of NLP in the healthcare sector can accurately provide clarity on unstructured medical data. It can also provide insights into understanding quality, refining techniques, and better results for patients. Two major architectures of AI are currently in use for NLP: Recurrent Neural Networks (RNNs) and Transformers. An important concept in NLP is the application of weight points to different words in a sentence (the more unique and clinically relevant the word, the more the weightage) and the application of a context vector to direct the AI model to its relevant conclusion. NLP can be used for recording medical data and converting doctors’ and nurses’ notes to patient records

3. Organized Datasets– These include laboratory values, costs, demographic data, clinical numerical data, and data that has been organized into tabular formats. Because this type of data has a fixed reference point (e.g., blood pressure greater than 140/90 mmHg will be considered hypertensive globally), simpler algorithms can be effective. Traditional machine learning algorithms help in interpreting this kind of data. There are different types of traditional machine learning architectures like Decision Trees, Random Forests, Logistic Regression, and Support Vector Machines. Organized data can be used for predictive analysis, research data collection, workflow analysis, financial analysis, and much more. The following flowchart depicts the workflow for creating an AI model suitable for healthcare purposes.

Challenges of applying AI in Healthcare

One of the most important considerations for any AI model to be successful is the quality, amount, and type of data used to train and validate the model. With medical data growing at a rapid rate, it is essential to use the latest and most accurate available data. It is also important to periodically upgrade the model based on new data. AI algorithms can only identify correlation. They do not prove causation.

Thus, it is important to verify whether the model employs correct methods to correlate the input and output. Many times, complex correlations predicted by the model are hard to interpret.

Currently, there is a lot of mystery, hype, misunderstanding, and scepticism as far as the relevance of AI in healthcare is considered. A major concern is the loss of jobs requiring repetitive work (clerical, billing, triaging, data compilation, etc.). Another concern is deskilling, which causes physicians to become overly reliant on AI, resulting in a decreased skillset of the physicians themselves. Automation bias, a phenomenon in which AI systems are considered to be superior to human intelligence, should be taken into consideration. Breach of privacy, data theft, and unethical use of data are all genuine challenges that are being addressed globally. 

This branch, though still in its infancy, if deployed in hospital workflows, keeping the interests of the patient in mind, has the potential to transform and revolutionize the healthcare sector.