In the fast-paced world of healthcare, data science plays a crucial role in improving patient outcomes and delivering quality care. One individual who has made significant contributions in this field is Ashwin Prakash, the Vice President of Clinical Data Science at Community Health Systems. With his extensive expertise and leadership skills, Ashwin has been instrumental in driving data analytics strategies and implementing innovative solutions to enhance the quality of care provided by Community Health Systems.
Ashwin brings a wealth of experience to his current role, having held key positions in renowned organizations within the healthcare industry. As the Chief Data Scientist at Post Acute Analytics, he played a pivotal role in developing cross-continuum post-acute care analytics solutions that helped providers improve patient satisfaction and decrease costs. Prior to that, Ashwin served as the Division Director of Data Science at HCA Healthcare, where he led data-driven initiatives and implemented predictive modeling to enhance clinical care delivery.
With an Executive Master of Business Administration degree from The University of Texas Dallas, Ashwin possesses a strong foundation in business strategy and leadership. Ashwin is also a Physician, has a PhD in computational biology from the University of Toledo College of Medicine and a Postdoctoral fellowship from Johns Hopkins School of Medicine. His educational background, combined with his technical skills in machine learning, genomics, and computational biology, enables him to bridge the gap between clinical expertise and data-driven decision-making.
1. Thank you for taking time to share your knowledge and experience about “Predictive Analytics in Healthcare” with our audience. Can you please start by telling the audience about yourself?
In my early years I had the opportunity to grow up in India and Africa, in both of those areas, there are significant health care accessibility issues owing to both cost and geography. It was clear to me that technology could play a huge role in breaking those barriers, so soon after medical school I decided to focus on helping to develop solutions for those issues. That decision took me away from bedside medicine and into the world of academia. I pursued a PhD at the University of Toledo Ohio, and then took up a fellowship at Johns Hopkins studying biomedical engineering. After several years of academic research primarily focused on Artificial Intelligence (AI) and machine learning (ML), I had the opportunity to lead Data Science for one of the divisions of HCA, and more recently transitioned from Hospital Corporation of America to Community Health Systems.
2. Can you tell our audience about Community Health Systems?
Community Health Systems (CHS) is one of the nation’s leading healthcare providers. CHS affiliates currently operate healthcare delivery systems in 46 distinct markets across 16 states with 80 acute care hospitals and more than 1,000 other sites of care, including physician practices, urgent care centers, freestanding emergency departments, occupational medicine clinics, imaging centers, cancer centers and ambulatory surgery centers.
Innovation in healthcare requires leadership and trust. A willingness to consider new ideas is essential in order to accelerate the pace of change and deliver better healthcare experiences for patients and clinicians. CHS is intentionally incorporating emerging technologies, artificial intelligence and other innovative solutions to advance patient care, to improve the work environment and to help address workforce challenges.
Our role in clinical data science within CHS is to enable the use of enterprise-wide data science and analytical capabilities to measure, benchmark and improve clinical results. We leverage machine learning and artificial intelligence to help improve outcomes by supporting clinical decision making and operations. This work enables “innovation from within” as well as the ability to partner with other organizations to co-develop new and innovative solutions.
4. How is predictive analytics being used in healthcare today, and what are some of its most impactful applications?
The traditional approach to adopting new innovation in clinical practice was to look at texts, journals, and expert opinion for evidence-based answers to clinical questions and establish best practices. While that approach will continue to remain important, we need to look no further than the recent COVID-19 pandemic to realize that other approaches will also be helpful. A tremendous amount of new data has been collected related to the care of patients, especially at large health systems such as CHS or national health systems in other countries. With predictive analytics and the use of ML and AI, we can bring inferences from this data straight to the bedside much faster than ever before. The new stepwise iterative approach could be described as Fail Fast/Learn Fast/Get Better. Financial systems across the globe have been using AI for years. We need to reach that level of sophistication in medicine.
This could truly revolutionize the way we approach patient care and management. One of the most impactful applications of predictive analytics is in disease prediction and prevention. These algorithms are designed to run in the background and look at data as they are entered in near real time. The system collects multimodal data such as vital signs, labs, medications ordered and administered, radiology etc, as well as clinical notes. This rich data can then be integrated to predict a patient’s acuity or risk of mortality for example. The information could be used as a decision aid in addition to conventional clinical algorithms such as SOFA scores. For patients predicted to have a high risk of mortality, the system could suggest discussing resuscitation or palliative care with the patient and family. On the other hand If the patient is stratified as low risk with improving parameters, the decision aid might suggest de-escalation of care or planning for discharge.
The true power of these algorithms come from the fact that while a human clinician may be able to consider and weigh 5 or 6 variables at a time as they make decisions, a machine can potentially take into account thousands of variables and their interactions at a time. So these algorithms could include variables of social determinants of health, or even historical likelihoods of outcomes across thousands of patients, in addition to clinical parameters specific to a patient in deciding the appropriate time and first site of discharge.This allows healthcare professionals to intervene early, implementing preventive measures and personalized treatment plans to mitigate the risk of readmissions or complications post discharge.
Additionally, predictive analytics plays a crucial role in optimizing healthcare operations and resource management. It can forecast patient volumes, predict disease outbreaks, and optimize staffing levels, ensuring that healthcare facilities can effectively meet patient demand. This leads to better resource allocation, reduced waiting times, and improved overall efficiency in healthcare delivery.
4. What are some of the challenges associated with implementing predictive analytics in healthcare, and how can they be overcome?
The opportunity for hospitals to use AI or predictive analytics to improve productivity and reduce costs is significant. However, there are a number of challenges that hospitals need to overcome in order to fully realize this potential. These challenges include a lack of technical expertise, data availability, and organizational readiness. Some of these challenges are:
- Lack of a clear business case. Many healthcare organizations struggle to articulate a clear business case for AI adoption. In some cases, this is because the potential benefits of AI are difficult to quantify. In other cases, the cost of AI adoption can be high, and organizations are unsure of how to recoup these costs.
- Data quality and availability. Data is a critical input for AI, but healthcare organizations often lack the data they need to support AI adoption. Data is often siloed across systems, making it difficult to aggregate and analyze. Furthermore, data quality can be poor, and privacy concerns can limit access to data.
- Lack of technical expertise. AI talent is in short supply, and healthcare organizations face challenges in attracting and retaining qualified AI professionals.
- Misaligned incentives. Healthcare organizations often have multiple and conflicting incentives, which can make it difficult to adopt AI in a way that maximizes value.
- Regulatory uncertainty. The regulatory environment for AI in healthcare is still evolving, which can make it difficult for organizations to adopt AI solutions.
- Lack of clinician and patient buy-in. Clinicians and their patients are often skeptical of AI and may be reluctant to use AI-enabled solutions. This can make it difficult to implement AI solutions successfully.
Addressing these challenges requires comprehensive change management strategies, including education, training, and continuous engagement with healthcare providers to build trust and confidence in predictive analytics. But despite these challenges, there are a number of factors that are likely to increase the adoption of AI in healthcare in the coming years. These include:
- Growing evidence of the benefits of AI. As more and more organizations adopt AI and share their experiences, the evidence of the benefits of AI is growing. This evidence will help to build trust in AI and encourage more organizations to adopt it.
- Increasing availability of data. The healthcare industry is increasingly investing in data collection and analytics capabilities, which will make it easier to adopt AI.
- Falling costs of AI. The cost of AI is falling, making it more affordable for healthcare organizations to adopt.
- Increased competition. As more and more companies develop AI-enabled solutions, the competition for healthcare organizations’ business will increase. This competition will force healthcare organizations to adopt AI in order to remain competitive.
- Increased adoption of cloud-based AI. Cloud-based AI is becoming increasingly popular in healthcare. Cloud-based AI offers a number of advantages, such as scalability, security, and ease of use.
5. How can predictive analytics contribute to personalized medicine and individualized treatment plans?
Predictive analytics is a powerful tool for advancing personalized medicine and tailoring treatment plans to individual patients. By analyzing vast amounts of patient data, including medical history, lifestyle factors, and treatment outcomes, predictive analytics can identify patterns and associations that are not readily apparent to humans.
This wealth of data allows our providers to better predict the likelihood of treatment success or failure for a particular patient, helping them make informed decisions about the most appropriate treatment options. For example, predictive analytics can be used to identify patients who are likely to not comply with recommendations for physical activity or nutrition and so respond unfavorably to a specific medication or treatment plan, reducing the trial and error approach often associated with treatment selection.
Additionally, predictive analytics can help identify patients at high risk of adverse events or complications, allowing healthcare providers to implement preventive measures or adjust treatment plans accordingly. This proactive approach improves patient safety and outcomes. CHS hospitals welcome about 54,000 babies into the world each year. Most of these deliveries are safe, successful, joyful events. But, childbirth is not without risk, and it is never just routine. An AI-based maternal-fetal early warning system and clinical decision support tool designed to create safer birthing experiences has been installed across CHS hospitals offering obstetrical services. The technology continuously monitors maternal vital signs, fetal heart rate, uterine contractions and labor progression to help identify, predict and alert birthing teams of potential issues for faster intervention during labor and delivery, so as to achieve the best outcome for both mother and baby.
In summary, predictive analytics enables healthcare professionals to move away from a one-size-fits-all approach and embrace personalized medicine, providing individualized treatment plans based on a patient’s unique characteristics and predicted response to interventions.
6. How accurate are predictive analytics models in healthcare, and what factors influence their accuracy?
The accuracy of predictive analytics models in healthcare can vary depending on several factors. One crucial factor is the quality and completeness of the data used to train the models. Healthcare data is complex and heterogeneous, which makes it hard to normalize data from all sources and ensure high quality. For example, many of the fields in our EMRs are free text entry, which leads to less than optimal training and input data for algorithms to function on, leading to less than optimal results.
Predictive analytics models require large and diverse datasets to identify meaningful patterns and associations. If the dataset used for training is small, incomplete, biased, or of poor quality, it can significantly impact the accuracy of the models. This is especially true when the problem is complex as this would need exceedingly more complex algorithms to be employed. Different algorithms have varying strengths and limitations, and the choice of the most appropriate algorithm depends on the specific use case. However a good rule of thumb to consider is that, the more complex the algorithm the more data it will need to reach optimal performance levels.
Image Source: Researchgate
It is important to note that while predictive analytics models can provide valuable insights and predictions, they are not infallible. Uncertainty and unforeseen factors can always influence outcomes, and healthcare professionals should use predictive analytics as a tool to aid decision-making rather than relying solely on its predictions.
While intuitively the concept of accuracy might seem like the right yardstick to measure and benchmark models, the truth is that it is imperfect at best. This is especially true when models are trying to predict events that are rare or relatively less frequent. For example, if we were to build a model that predicts in-hospital mortality and the raw mortality rate at a facility was 2%, all we would need is a model that predicts that every patient will be discharged alive, to achieve a 98% accuracy in prediction. So while accuracy is a simple and appealing measure, it provides limited utility in truly gauging the model’s utility. This underscores the importance of using complementary metrics like precision, recall etc. or even coming up with new metrics when assessing the effectiveness of newer technologies such as generative algorithms.
7. What impact can predictive analytics have on population health management and public health initiatives?
Predictive analytics can help to improve population health management and public health initiatives by providing insights that can be used to target interventions, track the spread of diseases, and allocate resources more effectively. By analyzing large datasets, including population demographics, health behaviors, and social determinants of health factors, predictive analytics can identify high-risk populations and target interventions to improve health outcomes. From a public health initiatives standpoint predictive analytics could improve allocation of healthcare resources, such as staffing and equipment through better prediction of demand for services. This helps ensure that healthcare services are accessible to those who need them the most and improves overall population health outcomes.
As evident from the COVID-19 experience a key application of predictive analytics is in disease surveillance and outbreak prediction. By analyzing data from various sources, such as emergency room visits, social media, traffic patterns, and other orthogonal data sources, predictive analytics can detect early warning signs of disease outbreaks or exacerbations. This allows public health authorities to respond quickly, implement targeted interventions, and prevent the spread of infectious diseases.
The Department of Health and Human Safety has prepared an inventory of planned Artificial Intelligence (AI) use cases in order to increase awareness of and cross-agency collaboration on AI initiatives (https://www.hhs.gov/sites/default/files/hhs-ai-use-cases-inventory.pdf). This is a significant step towards increasing the speed of AI adoption and scaling for applications in public health.
However, leveraging predictive analytics at the scale of population health is not limited to federal or state agencies alone, large health systems can also benefit from managing their patients through the continuum of care. At CHS through our physician practices, daily remote monitoring for patients with certain chronic conditions is enabling earlier interventions when needed and giving physicians more insight into how patients are doing between office visits. Enrolled patients with hypertension, heart failure and Type-2 diabetes use cellular-enabled devices to capture daily vital sign information which is transmitted to a virtual care team that provides day-to-day monitoring and additional care support when needed. This enables proactive interventions, such as lifestyle modifications or early screenings, to prevent disease progression and improve population health.
8. How can predictive analytics contribute to reducing healthcare costs and improving financial outcomes?
According to a recent paper It is estimated that AI adoption within the next five years using today’s technologies could result in savings of 5 to 10 percent of healthcare spending, or $200 billion to $360 billion annually in 2019 dollars, without sacrificing quality and access (Sahni Et.al. Jan 2023 https://www.nber.org/system/files/chapters/c14760/c14760.pdf).
The following are some examples of AI-enabled use cases that could help hospitals achieve these savings:
- Improve clinical operations (for example, capacity management, operating room optimization, Reducing emergency room wait times by using AI to predict patient arrival patterns and staffing levels.)
- Improving patient outcomes by using AI to identify patients at risk of decompensation or adverse outcomes.
- Better quality and safety (for example, condition deterioration management or adverse event detection)
- Automating administrative tasks such as coding and billing, which could free up clinical staff time to focus on patient care.
- Reducing supply chain costs by using AI to optimize inventory levels and predict demand.
- Improving clinical workflow by using AI to automate tasks such as scheduling.
- Improve claims management (for example, auto-adjudication, prior authorization or appeal denials)
- Improved Case management (for example, tailored care management, length of stay management or avoidable readmissions)
- Better provider relationship management (for example, network design or referral management)
(Source: https://www.nber.org/system/files/chapters/c14760/c14760.pdf)
The opportunity for hospitals to use AI to improve productivity and reduce costs is significant. However, there are a number of challenges that hospitals need to overcome in order to fully realize this potential. But recent market trends suggest that AI adoption in healthcare will likely continue to increase as organizations realize the potential for AI to improve outcomes and reduce costs.
In addition, large technology companies such as Google, Amazon, and Microsoft are increasingly investing in AI-enabled healthcare solutions. These companies have the resources and expertise to develop and deploy AI solutions at scale. Their entry into the healthcare market is likely to accelerate the adoption of AI in healthcare.
9. What are the potential ethical implications of using predictive analytics in healthcare?
The use of predictive analytics in healthcare raises important ethical and legal considerations that must be addressed to ensure responsible and ethical use of predictive models especially when they can influence care on the bedside.
A prominent ethical consideration is the potential for bias in predictive analytics models. If the data used to train these models is biased or incomplete, the predictions generated may also be biased, leading to disparities in healthcare delivery. It is crucial to address these biases to ensure representativeness, and regularly evaluate and validate predictive analytics models to minimize bias and ensure fair and equitable healthcare outcomes. Without this the models will reinforce these biases further and augment underlying inequalities.
A recent Pew research survey highlights how Americans feel about AI adoption in healthcare (https://www.pewresearch.org/science/2023/02/22/60-of-americans-would-be-uncomfortable-with-provider-relying-on-ai-in-their-own-health-care/). On one hand a large share of Americans think that the use of AI would reduce mistakes made by healthcare providers and improve outcomes, a significant proportion say that the problem of bias and inequity would get worse or stay the same. Another wide concern is that AI could make patient-provider relationships worse.
The security of their medical records is also a cause for concern shared widely. Predictive analytics relies on large amounts of personal health information, and protecting patient privacy is of utmost importance. Healthcare organizations must establish robust data governance policies, implement strict security measures, and comply with relevant regulations to safeguard sensitive patient data.
Three-quarters of Americans are greatly concerned about healthcare systems using these technologies to make clinical decisions and worry that the pace of adoption might be too fast. An important factor in addressing these concerns is developing responsible AI that needs to be invested in transparency and explainability. Providing explanations and justifications for predictions can help build trust and confidence in the use of predictive analytics in healthcare, both with clinicians as well as patients.
Another major concern is that the increasing adoption of AI and automation would lead to jobs being lost. Predictive analytics should be viewed as a tool to aid decision-making, and the expertise and experience of healthcare professionals should always be considered alongside predictive models. It is important to strike a balance between data-driven insights and the art of medicine.
8. How can predictive analytics in healthcare contribute to early disease detection and intervention?
As we have discussed before, early detection or better monitoring of chronic conditions can be facilitated by predictive analytics at a population health level. But in my opinion a key barrier in early detection of disease is related to access to primary care, and information related to nutrition and reliable guidance on healthy lifestyle choices. What is truly transformative about some of the new AI applications like ChatGPT, is that the barrier to accessing information is being broken down. While in the past medical information was only available through talking to a primary care provider or on the internet, people can now access this information through these chatbots. While there are certainly concerns of disinformation on these platforms, with appropriate levels of regulation and fine tuning these models could significantly lower the barrier for people to seek and answer their medical questions.
In a recent article the AMA underscored the important role applications like ChatGPT could play in delivery of medical education across disciplines (https://www.ama-assn.org/practice-management/digital/chatgpt-passed-usmle-what-does-it-mean-med-ed). Another study published in Plos Digital health evaluated ChatGPT’s performance on three exams that physicians need to take to obtain licensure (https://journals.plos.org/digitalhealth/article?id=10.1371/journal.pdig.0000198). The results suggest that these applications were able to perform at about a 60% accuracy out of the box without needing any specialized training. Other models since have brought that accuracy to over 90% just in the last few months. The AMA views this as potentially assisting in medical education of future doctors, physician assistants and nurses. This is very promising given the shortage of good quality primary care in many areas across the country. These technologies could potentially improve access and reduce costs, making it feasible for early detection and intervention of conditions.
9. What data sources are typically used in predictive analytics models in healthcare, and what challenges do they present?
There are a wide variety of data sources that can be used in predictive analytics models in healthcare. Some of the most common include:
- Electronic health records (EHRs)
- National databases like CMS MEDPAR, NHSN etc.
- Claims data
- Patient survey data like Press Ganey, Leapfrog etc.
- IoT from devices including operating robots, patient wearables
- Social media data
Each of these data sources has its own strengths and weaknesses. EHRs, for example, are a rich source of information about the longitudinal patient record, orders, diagnoses, treatments, and outcomes. However, they can also be incomplete or inaccurate due to having numerous free text input fields. They are also dogged by the problems of standardization and normalization as the first generation of EHRs were largely designed to transition paper based workflows into the digital space, without the intent to be mined for complex analytics. We are slowly seeing standards like FHIR which are gaining traction to make it easier to collate and use this data.
Claims data can provide information about patient utilization of healthcare services, but it does not typically include the complete information about the patient’s health history. Today there are significant barriers in being able to access and share data transparently between providers, payers and patients. Also fundamentally having health care insurance tied to employment creates discontinuity of the health record as patients transition from one employer to the next, needing to change providers who are in-network as a result. Variations across payers is a significant hurdle here.
Patient surveys can provide valuable insights into patient experiences and preferences, but they are based on sampling the larger population and can have their own biases and be difficult to generalize to the wider population.
Social media data can be a source of information about patient health behaviors and attitudes, but it can also be difficult to interpret and analyze. Genomics data can provide information about a patient’s genetic risk for certain diseases, but it is still in its early stages of development and is not yet widely available.
Wearable Devices and Remote Monitoring: Data from wearable devices, such as fitness trackers or blood glucose monitors, can provide real-time information about a patient’s health status. However, challenges include data accuracy, standardization, and ensuring data security and privacy. This data is also challenging because of proprietary formats and intellectual property issues that are barriers to easily integrate and share.
Addressing these challenges requires investments in data infrastructure, data governance policies, and collaboration between healthcare providers, technology vendors, and policymakers. Standardization efforts, data sharing agreements, and robust security measures are essential to overcome these challenges and unlock the full potential of predictive analytics in healthcare.
10. How can predictive analytics models be validated and evaluated in a healthcare setting?
Validating and evaluating predictive analytics models in a healthcare setting is crucial to ensure their accuracy, reliability, and generalizability. There are a number of factors that can be considered when evaluating predictive analytics models in a healthcare setting. These factors include:
- The clinical relevance of the model: Is the model able to predict outcomes that are important to patients and clinicians and have a meaningful impact.
- The performance of the model: As we have discussed before measuring how well the model performs in predicting real-world outcomes? Metrics such as accuracy, sensitivity, specificity, positive/negative predictive value, lead-time etc are useful for this.
- The interpretability of the model: Can clinicians or other end users understand how the model makes predictions and what inputs are resulting in that prediction.
- The ethical implications of the model: How might the model be used to make decisions about patient care and how it is integrated into the workflow.
In addition to the above principles of evaluation several scientific approaches can be used to validate these models:
- Retrospective Validation: Models are tested using historical data to assess their performance in predicting outcomes that have already occurred. This validation approach helps evaluate the accuracy of the models in a controlled setting. It is important to ensure that the historical data has not also been used to train the model in the first place. Typically historical data is split into a training set and a test set and the model performance is tested on the test set, this approach is called holdout validation.
- Cross-validation: This is an extension of the above idea in that the historical dataset is divided into multiple subsets of training and test, and models are trained and tested on different combinations of these subsets. Cross-validation helps assess the stability and robustness of the models by simulating their performance on different data partitions.
- Sensitivity analysis: This method involves testing the model’s performance under different assumptions about the data. This can help to identify the model’s strengths and weaknesses, and to determine how robust it is to changes in the data.
- Bias analysis: This approach is to use a variety of statistical tests to assess the model for bias. These tests can help to identify any patterns of bias in the model’s predictions that may not be apparent from a simple comparison of the model’s predictions to the actual outcomes.
- Prospective Validation: Models are implemented in real-time clinical practice, and their predictions are compared against actual outcomes. Prospective validation provides insights into the real-world performance of the models and their impact on patient outcomes.
- External Validation: Models are tested and benchmarked on independent datasets from different healthcare settings or populations to assess their generalizability. External validation helps determine if the models perform consistently across diverse populations and healthcare settings.
It is important to note that models also require constant validation and evaluation over time to ensure that they are performing at an optimal level and as intended. It is crucial to involve healthcare professionals, data scientists, and stakeholders in the validation and evaluation process to ensure a comprehensive continuous assessment. Continuous monitoring and refinement of models are necessary to adapt to evolving healthcare contexts and improve their performance over time.
Contact Details
