AI-Powered Diagnostics: Enhancing Accuracy & Speed In Healthcare

Recent advances in AI-assisted medical diagnostics tech are potentially revolutionary and developing at exponential pace. Medical diagnostics is inherently multi-modal. A physician examines patient data in multiple forms such as text, images, audio and more. Examples include radiology images, blood examination reports and patient interviews to cite a few. Experienced physicians consider all the above data points holistically to arrive at patient diagnosis and course of treatment.

One of the most impactful applications of AI in diagnosis is in the analysis of medical images. With AI algorithms trained on vast repositories of labelled data, a physician can leverage advanced imaging techniques to detect subtle abnormalities with unparalleled accuracy. Whether it's identifying early-stage tumours on an MRI or pinpointing fractures in an X-ray, AI streamlines the interpretation process, allowing for faster diagnosis and timely intervention. Medical imaging-based diagnostics have proved to have disease prediction accuracies exceeding 90 per cent in many cases.

From AI technology perspective, medical diagnostic models developed so far have been predominantly task-specific models which are based on deep learning and machine learning-based classifications of disease conditions. A chest X-ray interpretation model trained to detect tuberculosis is good at predicting TB with high accuracy and can beat a trained radiologist, but it will fail to detect other conditions such as pneumonia. It will also not be able to prepare a comprehensive report on the chest examination. This scenario is beginning to change.

Medical AI diagnostics is progressing from being task-specific towards being a generalist. Med-PaLM a large language model trained on medical domain was released by Google in late 2022 which was followed by Med-PALM-2 in quick succession in March 2023. Med-PaLM 2 demonstrated its matching human level of expertise in diagnostics by passing the US Medical Licensing Exam (USMLE) with 86.5 per cent. It also has generative AI capability to prepare long-form medical reports and summaries on par with expert physicians.

Another exciting development taking place is Medical AI being able to simultaneously consume image data such as X-rays and text data such Electronic Health Records and interpreting it holistically to provide diagnostic results. Google has released an early and experimental version of Med-PaLM-M where M indicates multi-modal generative AI.

Medical diagnostic AI is thus on a progression path from being task-specific to becoming medical generalist to ultimately becoming multi-modal getting closer to the holistic abilities of an expert physician. This creates an exciting possibility to make healthcare more accessible, equitable and reasonable to the global population in the future.

In a recent IDC survey of CEOs and CIOs on the likelihood and extent of business disruption caused by AI among various industries, life sciences and healthcare ranked among the highest at 44 per cent and 42 per cent respectively. Reasons are not difficult to find. At present, healthcare delivery is expensive, less accessible, and slower than it can be. Business leaders are seized of the opportunity and promise of medical AI and are becoming more confident of AI’s potential to innovate and disrupt medical diagnosis and healthcare delivery.

Factors important to successful integration of AI technologies in healthcare delivery are various. These range from trust, explainability, and regulation to the technology itself. Surveys have found that a significant number of patients have worries that their doctor may be relying too much on the use of AI in their diagnosis and treatment. AI output is often plagued by the black-box syndrome and explainability of the output becomes an important factor for doctors as well as patients. AI passing the regulatory guardrails is another important factor. In the US, more than 600 medical devices have been approved which have embedded task-specific diagnostic AI in them, but none have been approved so far which have generative AI of unimodal or multimodal variety built-in.

Given the extremely high degree of sensitivity associated with any ill-effects on human life however rare, medical AI technology will remain human-centric and tightly controlled. It will have to be monitored, approved and delivered by a responsible and qualified medical professional. A good way to view medical diagnostics will be as a doctor’s AI assistant, in other words a medical co-pilot whose job it is provide clinical decision support to the medical professional. This is an exciting field with huge potential for innovation and disruption in the not-too-distant future! 

Sanjay Deshmukh

Guest Author COO, Findability Sciences