When Deepmind’s AlphaGo defeated Lee Sedol, a top professional Go player, in the year 2016, deep learning (DL), a subset of machine learning (ML), became the buzz word instantly all over the world. After the demonstration of AlphaGo’s DL capabilities, businesses began to assess long term strategies to make serious commitment towards AI/ML implementations. Since then, the experimentations and use cases of implementation of AI/ML in US healthcare organizations have also increased exponentially.
As of 2020, AI/ML implementation is not only limited to traditional claims and clinical data in healthcare. It has expanded to use cases involving radiology images, human voices, handwritten text and spoken languages. The prediction of breast cancer, based on mammograms, to human levels of accuracy was a major milestone the healthcare domain achieved in image processing.
Recently, Amazon equipped Alexa users to integrate with Healthcare Payer systems and get their healthcare details on voice request. Natural language processing (NLP) has also demonstrated it’s capabilities through chat bots and sentiment/classification analysis based on user responses and reviews.
Although several use cases and results of AI/ML projects are promising and exciting, the execution of these projects is still a challenge.
The Steps to Success
An AI/ML project’s success depends on and is influenced by many internal and external factors. Some examples of the internal factors are technical competency, mindset/cultural change and the desire to adopt AI/ML project outcomes. Frequent changes in consumer behavior and complex adaptation of rapid technological advancements are examples of external factors.
Adapting to these changes begins with keeping an AI/ML project as simple as possible. It can be built upon and enhanced by adding new use case and empirical data.
Here are the important steps that we recommend to achieve success on your next AI/ML project:
Step 1. Define SMART Objectives
Each project should start with detailing out the expected results as S.M.A.R.T. objectives. S: specific, M: measurable, A: attainable, R: relevant, T: time-bound.
In the early phase of the project, some of the elements of your S.M.A.R.T. objectives may not be detailed enough. But efforts should be made to review them periodically and update them based on project progress and feedback. Here are a few examples:
Cost Predictions:
- Prediction of claim cost for next 1 year, after coronary artery bypass grafting (CABG) is performed on member
Quality of Care Predictions:
- Re-admission probability of a member within 30 days after discharge (following a specific procedure performed)
Resource Utilization Predictions:
- Future trends in number of calls received by customer services after a certain event like a festival gathering or a natural disaster
Step 2. Understand Your Data Better
Some of the key points to look for, at this stage, are data types and formats. For cost predictions, the transactional claims dataset may be adequate, but for member sentiment analysis, initial data may be present in a raw, linguistic form. Also, some datasets like claims have analysis limitation since the initial purpose of dataset generation is for reimbursement only. However, a combination of right claims data with other clinical & social data will be able to provide excellent cost predictions. So it is worth spending time and effort to better analyze every aspect of the datasets available, before moving to the next step.
Step 3. Identify Right Model
It would be evident by now to categorize use cases to know if you are looking for a prediction in values, probabilities, or classification. Prediction in values can be a request of the claim amount, auto-adjudication rate, etc. Probabilities are usually best suited when there are multiple expected outcomes, like in the case of a prediction of maximum reimbursement for rejected claim, before submitting an appeal.
The predictions would get aligned with the models to be used, regression vs classification. As research is expanding exponentially, new complex and use case specific algorithms are exhibiting the prediction accuracy like AdaBoost & XGBoost in Ensemble Techniques and Isolation Tree in Anomaly detection.
Experimentations with multiple machine learning algorithms would be required after deciding on the right category and predictions.
Step 4. Identify Right Performance Measures
Performance measures are quantifiable indicators to assess effectiveness and efficiency of a process or business. They must be stated distinctly and separately as compared to the performance measures of a related use case. There are metrics for predictive models to define how accurate it is to predict the right outcome relative to fact. On the other hand, for use cases, measures can directly or indirectly depend on model’s performance. For example, every 2% increase in prediction accuracy in claims dataset, there is a 4% increase in cost saving. It is often recommended to separately describe both measures, model performance & use case performance.
Step 5. Prepare Stakeholders
As you progress towards achieving the best AI/ML solution, it’s success ultimately depends on acceptance by stakeholders. It is recommended to have at least one representation from each stakeholder group in an AI/ML project since inception. This will eliminate the chances of last stage discoveries related to use case or medium by which the results will be delivered.
Once a use case and a corresponding solution are appropriately accepted, the next challenge would be to identify the medium through which the AI/ML results will be communicated to the stakeholder(s). The right medium of communication would always be crucial so that the largest population of stakeholders can comfortably receive and utilize the results.
Step 6. Continuous Improvement
Once organizations achieve the desired results in their initiatives, continuous efforts are required to achieve future goals. As more and more stakeholders adopt and witness positive outcomes, establishment of a feedback mechanism will improve the business values provided by AI/ML solutions. Improvement rate and stakeholder engagement in continuous improvement are key driving factors for success.
Parting Words
As new evidences of Reinforcement Learning (RL) and Transfer Learning (TL) application are emerging, it is just a matter of time before we see great achievement through new use cases in healthcare. An increase in complexity, due to continuous research and technological advancement and its integration with current IT infrastructure, can be deciding factor in the adoption of AI/ML by each business. There is still a long way to go for the healthcare domain as compared to others like retail, e-commerce, where whole businesses can revolve around AI/ML.
Leveraging AI/ML in Healthcare can be quite challenging as well as effort-intensive. But the advantages of implementing it are enormous and not to be missed in the 21st century.
Drop us a mail at info@nalashaa.com for a consultation on your AI/ML readiness!
Pankaj Kundu
Latest posts by Pankaj Kundu (see all)
- Here’s How You Can Implement Transparency in Coverage - July 22, 2021