Adoption of AI/ML can disrupt healthcare services – Express Healthcare

  • Lauren
  • June 11, 2021
  • Comments Off on Adoption of AI/ML can disrupt healthcare services – Express Healthcare

Read Article

Satish Kumar, CEO, Suparna Systems, Bangalore lists the many benefits of AI and ML and reiterates that rather than replacing health caregivers, these systems will lessen their administrative and documentation burden which will allow them to focus more on patient careL
Adoption of artificial intelligence (AI)/machine learning (ML) is going to disrupt healthcare services delivered, whether by Government or private hospitals and clinics. It will build better patients care and satisfaction and lower cost of delivery. It will also help in handling acute shortage of trained manpower in healthcare settings. This will also make healthcare services more accessible to rural regions.
Patients interact with the hospital for appointment booking, consultations, admissions, laboratory diagnostics, discharge, etc. In recent years several hospitals both private and publicly funded have adopted digital tools for automating the workflow and patient records. Hospitals have deployed electronic systems for appointments, admissions, diagnostics/tests, billings, insurance claims, pharmacies, etc. Record linkage connecting all these records for a particular patient remains a concern.
AI can enhance and augment patient interactions and workflow from primary care to the Emergency Room. It can be useful in EMS or paramedic service. AI-enabled symptom checkers can screen patients even before the patient visits the clinic or hospital. Chatbots can answer common questions regarding insurance eligibility, co-payments, etc and can help in appointment booking,
Screening chatbots can collect patient information and symptoms and can identify the correct specialty where the patient appointment can be given. This will help in reducing the waiting time at the hospital and bring patient satisfaction.
John Hopkins School of Medicine has used deep learning (a sub-branch of AI) to improve critical care in ophthalmology. A few years ago, researchers from Stanford University demonstrated that ML models can identify heart arrhythmias from ECG better than experts.
ML can build prediction models using Electronic Health Records (EHRs). EHRs are clinical data about the patient and have details of medical history, diagnosis, laboratory test results, medications prescribed, treatment plans, etc. EHR datasets are complex and heterogeneous. ML techniques of boosting and logistic regression can be used to predict serious conditions like heart disease etc. The ML models can be trained on the previous EHR records which contain such serious diseases as heart failure for which we want predictions.
To make it usable in real-life, ML algorithms have to be trained separately and exhaustively for each disease and medical condition. This will require a huge amount of data to train and validate, which is luckily available due to the widespread adoption of EHRs. Training the model is a costly and time-consuming process. The trained models need to be approved by the drug regulators before putting into actual practice. If the model has to be used across different countries, retraining and approval from country-specific regulators like FDA in the US or EMA for Europe will be required.
AI can also be used in risk stratification. It can classify the patients into high-risk and low-risk groups. In Emergency Room situations, it can help to predict the likelihood of mortality or admissions to ICU. There are research studies done to compare an artificial neural network (ANN) based prediction model and Acute Physiology and Chronic Health Evaluation II (Apache II) model for predicting ICU mortality using the same physiological variables. It was found that ANN-based prediction and classification model performance matched and in some aspects better than APACHE II.
For real-life use, the models have to be trusted by patients as well as doctors. Many health care providers are hesitant to adopt these models as they have no visibility of how the diagnosis decision has been made. These ML algorithms, particularly Deep Learning ones, behave as a black box. Which essentially means that you run it on your data, and you get results. You do not know the logic and reasoning behind your results. This creates a low confidence situation for patients and doctors. However, in recent years, a lot of research is getting done to make it transparent and explainable.
Besides clinical systems, AI/ML models can also help reduce the administrative workload of healthcare providers. AI can be used to auto-generate prescriptions from patient-doctor interactions. The auto-generated prescriptions can be run through doctors for approval and can be incorporated into the clinical workflow. These systems can also help in optimising hospital resources of beds, medical equipment, etc.
The models can be built cost-effectively as there are many high-quality open source systems and libraries available. For Deep Learning, TensorFlow and PyTorch can be used. For Natural Language Processing open-source libraries like AllenNLP or StanfordNLP can be used. The models will improve themselves and become better and more accurate the more they are used.
From the IT side, implementing and keeping them up to date will require building what is broadly known as MLOps. The organisation has to build continuous delivery and automation pipelines. Open source workflow orchestration software like Apache Airflow can be used for this.
These models are going to use the Personal Identification Information (PII), so ways to address privacy concerns must be implemented. Security also remains to be of paramount importance in these systems. Defense-in-depth principles must be used for building multi-layered secured architectures.
These AI algorithms are not going to replace the job of health caregivers, rather they are going to lessen their administrative and documentation burden which will allow them to focus more on patient care. One of the core reasons is algorithms do not understand trade-offs and the practice of patient care is an art and there are almost always trade-offs involved.
Also, AI only gives the probability, it does not give certainty. AI can show optimal/relevant choices and a physician can select one of them. The transformation of healthcare with adoption of these AI/ML models certainly is on the way.
 (Views are personal)