The field of AI is growing rapidly, and healthcare is poised for a transformative change. AI has the potential to greatly benefit the healthcare industry, but it must be embraced with cautious optimism. AI technology needs to be tailored to the unique challenges of healthcare in order to be successful.
Machine learning (ML) is already being used in a variety of ways in the healthcare industry, including:
Patient-directed technology: Wearable devices, AI chatbots, mental health and personal health apps, and genetic testing
Physician-directed technology: Medical imaging and pathology analysis, electronic health records (EHRs), and data analytics
Hospital-specific technology: Platforms like H2O.ai, Epic’s Deterioration Index
Clinical research: Genetic analysis, clinical trial design, and drug discovery
Telehealth: Telemedicine, disease management, and lifestyle management
ML/AI based technologies can help in monotonous, time consuming and laborious tasks. InnerEye is an AI system that can help doctors plan radiation therapy for head and neck and prostate cancer. It can do this much faster than traditional methods (up to 90% faster), which can save time and money for hospitals and patients and enables patients to receive cancer directed therapy sooner.
Expanding care/ services to underserved communities using ML/ AI have been documented. Data-driven artificial intelligence model for remote triage in the prehospital environment was successfully launched in South Korea. The model was trained on a dataset of over 100,000 patient records from the National Trauma Data Bank (NTDB). The model was able to predict the need for immediate hospitalization with an accuracy of over 90%. The authors also evaluated the model in a real-world setting using data from over 1,000 trauma patients. The model was able to accurately predict the need for immediate hospitalization in over 80% of cases.
AI is utilized in India in screening for cancer, osteoporosis prediction models and rehabilitation. BAGMO (blood bag monitoring) device addresses lack of blood availability in rural India.
The Nuance Dragon Ambient eXperience (DAX) is an AI‑powered, ambient clinical intelligence (ACI) solution that utilizes machine learning, natural language processing (NLP) to help in documentation. The physicians from healthcare systems who utilize this technology have reported 50% reduction in documentation time, 70% reduction in the feeling of burnout and majority of them feel that their quality of documentation has improved.
The Centers for Medicare & Medicaid Services (CMS) has approved Medicare reimbursement for the use of the Food and Drug Administration (FDA)-approved AI algorithm IDx-DR. IDx-DR is a system that can detect more-than-mild diabetic retinopathy with 87% sensitivity and 90% specificity.
A Novant Health facility used a predictive model called the Epic Deterioration Index (DI) to identify patients who were at risk of getting worse more quickly. This helped the hospital staff to provide better care to these patients, which reduced the number of deaths by 22% and saved an estimated 153 lives over 11 months.
This has been a complex issue especially when dealing with healthcare and sensitive patient data. Risks can be mitigated by using de-identified data, when possible, data encryption and regularly auditing. Data safety can be maintained by partnering with secure cloud computing provider and having only authorized personnel access to the data.
The process of clinical trial matching is complex. Lot of information needs to be processed – inclusion and exclusion criteria, performance status, biological markers, next-generation sequencing (NGS) results, etc. Yuma regional medical center has partnered with industry to open precision oncology clinical trials in real time. Data ingested was from electronic medical records through EMR integration or clinic, pathology reports and NGS results were obtained from the labs.
For the successful implementation of the program both methodologies were used- large amounts of data from the NGS labs were used to train a standard NLP model using annotated data. Also, we leveraged domain experts to validate the unstructured records from each patient which was obtained through the electronic medical record. There was a large amount of relevant data available from these sources.
From the high-level patient matches, another domain expert manually evaluated the data to confirm that the match was appropriate. Subsequently, the information was sent to the site about a potential clinical trial for the patient.
This led to successful enrollment of patients in relevant oncology clinical trials at our institution. We published this as an abstract at the American Society of Clinical Oncology conference.
The Swedish group has shown that AI-supported mammography screening when compared with standard double reading has similar cancer detection rate. AI has shown good accuracy and sensitivity in identifying imaging abnormalities. There remains the need to balance enhanced detection Vs overdiagnosis which AI tools can help with. These technologies will be an ally to the radiologists and not eliminate the role of radiologists. They have the potential to enhance the workflow by identifying normal digital breast tomosynthesis screening examinations and decreasing the number of examinations that require radiologist interpretation.
The company Freenome is using AI to detect weak signals in blood to diagnose cancer in early stages. The company analyzes fragments of DNA, RNA, and proteins in blood plasma. Its multiomics platform identifies key biological signals. The company deploys ML to understand additive signatures for detecting cancer early.
Personalized treatment recommendations based on patient data can improve patient outcomes, be able to provide timely care and reduce healthcare costs. A new AI system developed at Johns Hopkins can detect sepsis earlier. Traditional methods of detecting sepsis can be slow and inaccurate, but the new AI system is able to identify patients at risk much earlier. This means that doctors can start treatment sooner, which can improve a patient's chances of survival. It has the potential of reducing the risk of death by 20%. The model sifts through electronic medical records and clinical notes to identify patients who are at high risk of developing serious complications.
Google DeepMind has developed a new AI system that can predict acute kidney injury (AKI), a serious kidney condition, up to two days before it happens. This means that doctors can start treating AKI sooner, which could save lives.
H2O.ai's AI system can look at all the data from a hospital or healthcare system to predict which patients are most likely to need to be transferred to the ICU or to get an infection while they are in the hospital.
A new drug designed by artificial intelligence is now being tested in people. The drug is called INS018_055, designed by Insilico Medicine and it is being tested as a treatment for idiopathic pulmonary fibrosis (IPF), a chronic lung disease. The drug has completed phase 1 human safety studies and entered multi-regional Phase II clinical trials. The study is being conducted in the US and China and it will involve 60 people with IPF. The study will last 12 weeks and it will assess the safety, tolerability, and effectiveness of the drug. If the drug is successful in this study, it could be the first AI-designed drug to be approved for use in patients. This would be a major milestone in the development of AI-powered drug discovery, and it could lead to more effective and personalized treatments for a wide range of diseases.
The challenges that need to be addressed are regulations, access to data and ensuring that the data is truly representative of the population, thereby reducing biases. Other barriers include high costs, ethical concerns and interpretability.
Risk mitigation strategies should involve use of algorithms that are fairer and more transparent. There needs to be better understanding among researchers around the guidelines. Regulatory agencies are struggling to keep pace with the rapid evolution of AI. As technology advances, the cost of developing and deploying AI models will continue to go down. Also, legislature and government funding can help with AI based research. AI systems need to be more interpretable, i.e., we should have better data visualization of the AI predictions and the models should be able to explain their decisions.
AI systems need to be transparent and explainable. One needs to understand how the system makes these decisions and be able to identify any potential biases. Black box AI systems often involve complex algorithms such as deep neural networks. They can be effective in solving problems but difficult to understand. There is a potential to have biases built in in those systems especially if the system is trained on biased data.
The ethical concerns regarding deploying ML/AI in healthcare revolve around trust, consistency and explainability. Work needs to be done to minimize bias and promote fairness in algorithms. We want to ensure that the data on which ML/AI algorithms are trained are free of bias as it can lead to adverse or unfair outcomes for the minority population. ML/AI algorithms need to be transparent about how these algorithms work and be able to provide information about how data is being used. There also needs to be accountability by possibly having human review. These processes will help in improving patient outcomes and building trust with the public.
Interpretability and explainability of AI ML/AI models are crucial for healthcare decisions and for building trust. ML/ AI algorithms need to be interpretable, consistent and trustworthy. The AI techniques used need to be transparent and explainable. As mentioned earlier, having human review of the decisions made by ML/AI algorithms can ensure fairness and help build trust. Also sharing data with patients about how these algorithms work can ensure buy-in from all stakeholders.
The black box problem is the lack of understanding by which a ML/AI algorithm arrived at a particular conclusion is a matter of concern by healthcare organizations, public and regulatory personnel. For FDA approvals for medical devices, it is important to demonstrate that the device is effective, safe and has demonstrated success by enrolling sufficient patients in clinical trials. Other key factors include data privacy, security and ongoing compliance to ensure that the solution continues to comply with the latest regulations and standards.
One of the biggest challenges in the use of ML/AI in healthcare is that regulations are not keeping up with the rapid pace of innovation. Additionally, healthcare systems are complex organizations with multiple stakeholders and increasing regulations around data privacy and security.
Despite these challenges, ML/AI has the potential to revolutionize healthcare by improving patient outcomes, reducing costs, and providing more timely access to care, especially for underserved communities.