Vishal Singal

For a child with autism spectrum disorder (ASD), effective care coordination across the required medical, physical rehabilitation, counseling, occupational, and speech providers is challenging, especially when the aim is to provide individualized care—at scale across many patients.

Genial Care, a healthcare company focusing on personalized interventions for children with ASD, provides a successful example of how generative artificial intelligence (GenAI) can solve this problem. GenAI monitors sessions with children and their families, synthesizes session records, improves the records’ quality, tracks each child’s progress, and shares this information with providers and the child’s family via an app-based interface. Plus, the company reduced the manual transcription time from approximately two hours per day to three minutes per session, allowing care providers more time to dedicate to their patients.

However, this is one of only a few examples of a life sciences or healthcare company breaking out of GenAI pilot purgatory. How did they do it? They focused on a practical, human-centered use case that easily integrated into daily workflows and could be efficiently scaled.

Why is the life sciences industry stuck in GenAI “pilotitis"? Several reasons come to mind:

Regulatory constraints around patient data

Risk aversion, stemming from a smaller margin of error related to life-altering decisions, which can be exacerbated by mistrust due to AI hallucinations and viewing AI as a “black box”

Lack of an enterprise-wide AI strategy and organizational readiness, often resulting in siloed efforts, wasted resources, and declining innovation efforts

Limited input from stakeholders, resulting in solutions that don’t solve real problems or fit into existing workflows

Choice of complex use cases with long implementation times and limited impact, resulting in slowed momentum

The light bulb moment: GenAI's potential to illuminate healthcare

The incandescent light bulb’s journey to widespread adoption perfectly illustrates the hockey stick-shaped trajectory of transformative technologies like GenAI. Initial development, which spanned decades from 1835 to Edison's commercialization in 1880, was followed by decades of infrastructure development (power plants, transmission lines) necessary for widespread use. This gradual ascent reflects the “entropy effect”—the disruption to established norms that slows adoption.

Similarly, AI's transformative potential in life sciences, while offering increased speed and accuracy in tasks like synthesizing unstructured data, continues to face resistance, leaving it in the slow ascent phase. To accelerate AI adoption, we can learn from other sectors already leveraging AI at scale.

From auto racing to healthcare: The transformative power of GenAI

Like healthcare, other industries suffer from an overwhelming amount of data. For example, in the world of automotive racing, although a race lasts hours, fans often want a quick summary of the highlights and outcomes. In this setting, a GenAI application was developed to produce a two-minute summary podcast in any language from two hours of race commentary.

 LLM enrichment using driver data and future race events further set the system up for the next race.

In healthcare, this same approach could be used to:

Summarize patients’ entire health histories

Synthesize healthcare professional (HCP) feedback

Suggest next best steps for HCP engagement

Predict future health outcomes and treatment-related behaviors

Streamline benefit verification by payers

We also often want to find specific information from within messy notes or audio or video files. Globant’s Advance Video Search does just that.

 It uses information from a text- or image-based search to prompt GenAI to search vast amounts of video. It can find specific clips, images, and frames within the video files.

Imagine if that same precise search functionality could be used to find:

Personalized, compliant content from a company’s website or patient support site based on a patient’s natural language search terms

Specific patient insights from unstructured clinical notes

For example, we’ve implemented conversational AI solutions in production for multiple life sciences clients where they are able to find insights and answers from large libraries of analyses and notes. Tasks that previously took many hours of painstaking manual research can now happen in minutes. Subtle, yet valuable insights that may have been overlooked before are being brought to the surface.

The Genial Care example mentioned above demonstrates that we are also seeing examples of scalable GenAI use emerging in life sciences. It could also be applied to more than ASD, such as complex rare diseases or patients with comorbid chronic diseases who require care from many different providers.

In addition, Predictive Oncology’s PEDAL platform harnesses AI to make predictions about drug responses in oncology.

 Combining AI, data about drug responses, and a biobank of >150,000 tumor samples from 137 tumor types, the platform uses machine learning algorithms to make predictions about the effect of different drugs on tumors, with an accuracy of 92%. Although the company is using these for drug discovery and repurposing, this type of functionality has implications for clinical care, where the right drug could be identified for the right patient, at the right time.

Another example from a Beghou client involved predictive analytics leveraging 20+ variables into five easily discernable attributes to help drive field alerts. The alerts and suggestions were highly relevant, with the field team taking action on over 75% of them, much higher than industry benchmarks. One alert identified a rural HCP with a potential rare disease patient, which caused their team to prioritize that HCP and ultimately lead to a new patient start.

The AI-generated “digital brain” from Orby is changing neurorehabilitation and pain management for patients who have lost their motor skills due to spinal injury, stroke, or other conditions.

 By mapping neurostimulation to affected brain areas, the company’s able to tap into specific disordered functions. Applications for mental health issues come to mind, where targeted treatment could be delivered.

Within this realm of exciting possibilities, some limitations of GenAI also need to be acknowledged, or what it is not:

These misconceptions feed into the entropy effect, mistrust, and reluctance for adoption. They also highlight the important roles of humans to ensure relevant, accurate, and trusted results, by guiding its outputs; providing oversight to mitigate bias in the data inputs; reviewing the results for accuracy; and giving feedback to improve machine learning processes.

What does this look like in practice? Recommendations could include the following:

Identify unmet needs with the organizational readiness to solve them:

 Find issues that are not currently being optimally met by existing systems and processes where the transition to AI would make a meaningful impact to end users (e.g., commentary summaries, synthesis of unstructured notes, or live audio).

Prioritize low-complexity, high-impact use cases with a focus on scalability:

 Review your high-impact use cases for the level of complexity to implement. Tasks that already have proven AI solutions (i.e., are low complexity), such as tagging and summarizing unstructured data to better understand customer sentiment, result in quick wins that maintain momentum.

 Establish a powerful, compliant infrastructure of data, technology, and governance to ensure long-term scalability.

Center people in the development and implementation: 

Field an end user panel to provide input and feedback throughout design, development, and review, and leverage your team’s expertise to drive context, accuracy, and judgment.

Conduct pilot testing for early exposure and ongoing refinement:

 Allow users to provide feedback during the build to generate excitement, foster trust, and lay the foundation for enterprise-wide adoption and expansion.

Incorporate change management processes: 

Make sure your entire organization is aligned and on board regarding the uses and benefits being pursued with AI, so that, when you go live, your organization speaks with a unified voice and makes a splash.

We believe that stalled or failed GenAI pilots can become a thing of the past in life sciences. Although data-related factors such as volume, structure, and compliance have proven to be problematic for extracting valuable insights and personalizing care, we’ve found that they can be effectively addressed using a human-centered approach and enterprise-wide strategies that drive scalable AI solutions.

Dan Schulman and Vishal Singal are both partners at Beghou Consulting.

1. Google and Formula E Break World Record for Most Participants in Generative AI Hackathon. 

https://www.fiaformulae.com/en/news/504402

2. Globant Releases Innovative Advanced Video Search Tool Leveraging Google Cloud’s Gemini Models. 

https://www.globant.com/news/globant-releases-advanced-video-search-tool-avs

https://predictive-oncology.com/poc-report-page/

Articles

By embracing practical, human-centered use cases, life sciences companies are breaking free from GenAI “pilotitis” and demonstrating how scalable AI solutions can help transform patient care and provider efficiency.

Innovation with GenAI: Lessons from Life Sciences and Beyond

What if we could flip the patient data iceberg to expose what is typically hidden from sight—the full narrative of a patient’s healthcare journey? In the 10% of the iceberg we see above the surface are the traditional data sources such as claims and diagnostic and procedure codes, which give only us a snapshot of the patient. It is the vast base underwater that provides the rich detail we need. When paired with advanced analytics, this data can help us reduce time to diagnosis, understand the “why” of patient behaviors and decisions, and identify patient cohorts for enhanced patient engagement and interventions.

The industry’s current approaches to patient data analytics focus on structured data (e.g., height, weight, standardized test results, diagnosis codes, medications, stages of disease diagnosis). Although this data helps capture the patient’s journey, it does not represent 100% of the patient’s interactions with the healthcare system. In fact, only 20% of patient data is captured in structured form, with the remaining 80% comprising unstructured data (e.g., physician notes, progress notes, infusion summaries, lab reports). Furthermore, according to RBC Capital Markets, the compound annual growth rate of data for healthcare will reach 36%, and there will be nearly 5,000 digital device interactions daily per person by 2025.

 All this will significantly increase the amount of unstructured data.

The combination of increasing variance, volume, and velocity of patient data overall and complex unstructured data, which is not easily codified, has limited the industry’s ability to understand the full patient journey. For life science companies developing niche and targeted therapies, the resulting lack of insight creates challenges for identifying the right patients and intervening at the right moments to support treatment journeys.

Moving the patient health paradigm from 'what' to 'why'

Structured data primarily answers the “what” of patient health—clinical diagnoses, current disease status, treatment initiation, and discontinuation. Unstructured data provides the context around their health—the “why” —a comprehensive longitudinal view of the patient, a road map to their physicians’ thought processes, and the ability to understand the clinical and socioeconomic aspects of patients’ reactions to therapies and diagnoses.

For example, clinical notes capture nuanced observations about patient behavior, drug responses, and lifestyle factors; laboratory narratives and genomic data reveal deeper biological patterns and potential therapeutic targets; and detailed pathology reports provide context around disease progression. Advanced natural language processing (NLP) algorithms can extract meaningful insights from these unstructured sources, converting them into analyzable data points. Combining these with structured data sources in a way that is compliant with privacy requirements can generate a more comprehensive patient data set.

Accelerating diagnoses using unstructured data

A delayed rare disease diagnosis can be not only mentally and emotionally draining for patients, but also financially draining. According to a study by The EveryLife Foundation for Rare Diseases, the “diagnostic odyssey” for a patient with a rare disease can result in an average of more than $220,000 in medical costs and lost income.

 Each year saved in diagnosis directly impacts patients. We have seen in our work with life sciences companies that it can take up to eight years and 15 (or more) clinical encounters for patients to secure a diagnosis.

Therefore, finding new ways to speed diagnoses could significantly improve patient outcomes and finances. Utilizing artificial intelligence (AI) on unstructured data can fill gaps in a patient’s diagnostic journey.

For example, relatively simple documentation of the progression of fatigue in primary care notes, when combined with other longitudinal elements of the patient journey, can be an early indicator of rare pediatric diseases with a genetic predisposition that might be asymptomatic at birth. NLP-based analysis of physician notes could flag the need for appropriate specialist referrals for further diagnostic tests, ultimately reducing the time to diagnosis.

For other diseases, the use of AI to analyze patient data could help reduce misdiagnosis or underdiagnosis. The biggest challenge for one of Beghou’s clients with a rare disease drug was finding patients and initiating treatment. Misdiagnosis often occurred because of the multiple levels of symptoms, most of which were mild in the initial disease phase, and considerable delays in the confirmatory genetic test for most patients. Therefore, instead of using only claims data, which identified very few patients, we identified patterns of symptoms from patient notes. These symptoms ranged from simple skin spotting, vertigo-like symptoms, decreased ability to sweat, gastrointestinal issues, kidney-related issues, and irregular heartbeat. Using AI models, a substantial number of patients, each with a bundle of these symptoms, were identified within regional hotspots, which enabled focused outreach and facilitated diagnosis.

Enhancing engagement and personalizing interventions

Detailed patient profiles can also be developed using advanced analytic techniques such as machine learning and predictive modeling on clinical data (e.g., electronic health records, lab results, treatment responses), behavioral patterns (e.g., medication adherence, appointment attendance, portal usage), and socioeconomic factors (e.g., education, income, access to care) from both structured and unstructured data.

These profiles can be used to identify distinct patient microsegments with unique needs and preferences based on subtle patterns in disease progression; objective and subjective treatment effectiveness; predictions of potential complications; engagement preferences and compliance tendencies; and barriers to care and access to resources.

Patient microsegments facilitate the development of more effective, personalized intervention strategies and patient-centric, outcome-driven engagement plans, such as simplified medication schedules for patients with complex situations, digital health tools for tech-savvy individuals, and high-touch support programs for those requiring additional assistance.

For example, for patient cohorts predicted to be less adherent to their treatment, providers can understand underlying factors such as adverse effects, social determinants of health, or misunderstandings about the importance of treatment. Appropriate intervention strategies could include adverse effect management, reminder notifications, digital app-based prompts, and education about the treatment itself.

Improving insight quality and reducing cycle times with GenAI: A real-world example

We used data science and AI to improve the quality of insights by 25% and reduce intervention implementation cycle times by 20% for a specialty client. The GenAI program enabled and scaled AI infrastructure across the enterprise and functions to enable timely interactions with the healthcare providers and patients. Raw data from various unstructured data sources were enriched with the client’s specific ontologies and medical vocabulary to ensure the model could address ill-framed or abstract questions and integrate insights using retrieval-augmented generation. The latter allows the model to handle large amounts of data with relative ease and provide contextually relevant responses for the brand from all data sources, with a focus on usability via conversational AI, executive summaries, insights dashboards, and semantic search.

In some ways, claims data sets represent only the tip of the iceberg. With rapid technological advancements, especially with GenAI, we are on the brink of a large-scale transformation where combining structured and unstructured data sets will truly revolutionize how we understand patients and their journeys, engage and provide meaningful support, and impact patient outcomes.

To turn this potential into reality, we must radically change how healthcare stakeholders use unstructured data. This relies not only on tech stack, data storage, AI, and machine learning advancements, but also on data security and structural and change management. Cross-functional collaboration with clinicians and data scientists is fundamental to:

Understanding the nuances of medical ontologies.

Normalizing advanced AI algorithm use across healthcare systems.

Optimizing storage across disparate systems using common data models.

Refining definitions in metadata to evolve with advancements in medicine.


Instituting change management across functions for data democratization.

With a holistic, collaborative approach to designing richer data ecosystems and implementing advanced technologies, we can gain an in-depth understanding of health-related behaviors and decisions. Interventions centered around the patient then become possible, providing benefits for patients, physicians, and life science companies alike. Are we ready?

Sonam Dubey and Vishal Singal are partners at Beghou Consulting.

1. RBC Capital Markets. The Healthcare Data Explosion. 2020. 

https://www.rbccm.com/en/gib/healthcare/episode/the_healthcare_data_explosion

2. EveryLife Foundation for Rare Diseases. The Cost of Delayed Diagnosis in Rare Disease: A Health Economic Study. 

https://everylifefoundation.org/wp-content/uploads/2023/09/CombinedInfo_PDF.pdf

Filling the gaps in the patient journey within an increasingly complex data environment. 