This blog is a summary of a podcast Medable sponsored that occurred on June 4, 2025 and was hosted by Emerj.
You can find it here, https://podcast.emerj.com/the-evolving-role-of-ai-in-modernizing-clinical-trials-with-xiong-liu-of-novartis
Artificial intelligence is shifting from promise to practice in biopharma, and few domains feel the impact more than clinical development. In a recent conversation on the AI in Business podcast, Zhong Lu, director of data science and AI at Novartis, outlined how AI is evolving the clinical trials process end to end.
His perspective centered on building internal platforms that unite structured and unstructured data, deploying document intelligence to remove process friction, and using responsible generative AI to support scientific decision making without compromising compliance. While the regulatory bar remains high and patient trust is paramount, Novartis’s approach shows how targeted applications of AI can streamline data operations, strengthen regulatory readiness, and make both traditional and decentralized trials faster and more resilient.
Clinical trials are moving from digitization to intelligence
A year ago, many life sciences organizations were still focused on basic digitization, including eConsent, electronic data capture, and moving off paper. Today, the conversation has progressed to connecting data and making it intelligent. That means orchestrating information across systems so teams can drive decisions in real time, particularly in settings where accuracy, safety, and auditability are non-negotiable.
At Novartis, AI is being applied to optimize protocol design, predict site-level risks, and surface early safety signals. The aim is not to replace established processes but to augment them with timely, explainable insights that reduce rework, shorten cycles, and improve confidence. This shift marks a meaningful evolution from tools to orchestration, where value comes from how systems work together rather than from any single application.
Internal platforms unify structured and unstructured data
Clinical programs generate vast volumes of structured data, including imaging, labs, EDC, and EHR extracts, alongside an even larger body of unstructured content, such as protocols, clinical notes, adverse event narratives, and correspondence. Much of the essential context behind decisions lives in these documents. Novartis is investing in internal tools that integrate both worlds, pairing data architectures that can be queried in natural language with document intelligence that extracts tables, figures, and narratives from PDFs and other formats at scale. By connecting these layers, teams can retrieve evidence quickly, trace how insights were generated, and repurpose verified content into downstream deliverables. The platform approach also lays the groundwork for proactive readiness, assessing study health continuously and flagging issues before they become delays.
Decentralized trials become a catalyst for efficiency and access
Decentralized clinical trials are no longer a thought experiment. Enabled by telemedicine, wearables, and remote monitoring, DCTs reduce site burden and expand access for patients who might otherwise be excluded due to geography or time constraints. AI strengthens these models in two ways. First, it enhances outreach and enrollment by precision-targeting potential participants through patient portals, registries, and compliant digital channels, while improving two-way communication with investigators and sponsors. Second, it harmonizes data generated across homes, clinics, and labs, providing sponsors with near real-time visibility into safety and performance. In practice, most DCTs are hybrid, combining site visits with remote touchpoints, but the operational gains are tangible, including fewer missed visits, better data completeness, and more inclusive cohorts.
Federated learning and digital twins reduce friction without sharing raw data
One of the most promising areas of innovation is privacy-preserving analytics. Federated learning allows institutions to train models on local data and share only model parameters, not patient-level records, back to a central server for aggregation and optimization. Iterating this process yields a consensus model that benefits from multi-site diversity without moving sensitive data. The approach proved its value during COVID-19, where hospitals collaboratively predicted oxygen therapy needs using local vitals and imaging while maintaining privacy. In parallel, digital twins, which are algorithmic representations built from trial and real-world data, support external control arms, dose optimization, and scenario testing. These techniques do not eliminate the need for clinical evidence, but they can reduce control-arm burden, sharpen eligibility criteria, and make study designs more robust from the outset.
Responsible and compliant by design
AI in clinical development lives under the strictest scrutiny. Novartis’s approach embeds governance, explainability, and security from day one. Document intelligence pipelines are auditable. Model lineage, versioning, and access controls are enforced. Sensitive data are de-identified where appropriate, and privacy-preserving methods such as federated learning are prioritized to minimize exposure. Equally important, roles are clearly delineated. Not every researcher needs direct access to patient data, and specialized teams manage consent, privacy, and partner engagement to maintain compliance across jurisdictions. Treating trust as a first-class design criterion ensures that innovations are deployable, sustainable, and ready for inspection by regulators and internal quality teams.
From reactive operations to proactive readiness
The operating model is evolving from reactive to proactive. Instead of waiting for audit findings or enrollment bottlenecks, AI services continuously assess study readiness and recommend next steps. Protocols benefit from lessons learned across prior trials and real-world outcomes, enabling teams to anticipate risks and automate elements of submission drafting. Anomaly detection highlights outliers in operational metrics or safety signals, prompting early interventions. Document intelligence accelerates medical writing by grounding generative outputs in verified data, which reduces cycle time without sacrificing accuracy. Across these use cases, humans remain squarely in the loop. The objective is to augment clinical, regulatory, and legal teams with tools that surface the right information at the right time.
Accessibility and scale determine enterprise impact
For AI to deliver business outcomes, it must be accessible to the people doing the work and scalable across programs. Novartis is integrating AI into the systems clinicians, monitors, writers, and quality teams already use, adding clear explanations and source citations to build trust. Platform thinking replaces one-off use cases. Common services handle identity, security, audit trails, model governance, and monitoring, so new applications can be onboarded faster and with less risk. Success is measured in operational terms, such as fewer protocol amendments, faster safety reviews, improved audit readiness, and stronger enrollment and retention, so value is visible to leadership and reinvestment is justified.
How leaders can get started
For organizations earlier in the journey, the path forward is pragmatic. Start with the opportunities that remove process bottlenecks without touching high-risk decisions, such as document intake and summarization, search across validated repositories, and operational dashboards grounded in traceable data. Build a data fabric that connects core clinical systems to a secure document intelligence layer, and standardize on model governance practices before scaling. Pilot decentralized elements where they make enrollment and retention meaningfully better, and use privacy-preserving methods to collaborate with partners without moving raw data. Above all, align each initiative to a specific business outcome, such as cycle-time reduction, compliance readiness, or enrollment lift, so momentum compounds and stakeholder trust builds with every release.
Conclusion
AI is not a silver bullet for the cost, complexity, and risk of clinical development. However, when it is deployed with intention, it becomes a durable advantage. Novartis’s approach, integrating structured and unstructured data on an internal platform, applying document intelligence to eliminate friction, and using responsible generative AI to support scientific teams, demonstrates how targeted investments can lift quality and speed at the same time.
As decentralized models mature and privacy-preserving analytics become commonplace, sponsors will rely on proactive readiness rather than reactive remediation, and clinical operations will feel less like orchestration by exception. The result is a more inclusive, efficient, and compliant path from study design to submission, one where patients, investigators, and regulators share a clearer view of progress, and where scientific decisions are grounded in data that are connected, explainable, and ready for scrutiny.
.webp)
