We connected with Tina in her office in Germany, where she took time from her busy schedule to discuss the state of AI, innovation, and creating scalable value in MedTech and Life Sciences.

‍

Loft: AI has significant potential for imaging and image processing, particularly in fields like radiology. However, challenges persist in training these algorithms due to the difficulty in data acquisition or the flaws within datasets. Considering the aspirational goal in cancer treatment, where an algorithm could potentially prescreen 100% of a dataset and only require a physician to review 1%, leading to more frequent screenings and potentially saving lives, the questions that arise are critical. Can we trust the algorithms sufficiently? What are the trade-offs between fidelity and quality, and between fidelity and magnitude in these AI systems? How do these factors influence AI's overall reliability and effectiveness in medical imaging?

‍

Tina: AI holds substantial potential in healthcare, particularly in improving clinical, operational, and financial outcomes. The greatest challenge lies in realizing clinical value due to stringent regulations concerning patient safety and quality. Operationally, AI can enhance efficiency, such as by analyzing log files from medical instruments to optimize scanner throughput or determining the most effective resource allocation for treatments like stenting. Clinically, while AI can assist radiologists by highlighting potential oversights in scans, implementing these tools as medical devices involves complex regulatory approvals. However, advancements in AI, like HeartFlow’s use of cardiac CT images to assess the necessity of catheter-based procedures, demonstrate significant benefits. This technology not only improves patient outcomes by reducing unnecessary invasive treatments but also cuts down on medical costs and reallocates resources to more urgent cases. Achieving high accuracy in clinical AI applications is challenging, requiring extensive and diverse data sets. The advent of Generative AI (LLMs) and the use of synthetic data has somewhat eased these challenges by reducing the amount of real data needed and improving algorithm robustness. Synthetic data isn't a substitute for real data but complements it by preparing the AI to handle real-world scenarios it might not have previously encountered. Despite these advances, there are still pitfalls. AI systems can't guarantee uniform precision across all patient groups, and there are variations based on different geographical, cultural, and environmental contexts. Continuous learning and transparency about the data used for training and the intended use of AI systems are crucial. By continually refining AI algorithms based on real-world feedback, we can enhance their effectiveness and safety, fostering trust and broadening their applicability in healthcare.

‍

This technology not only improves patient outcomes by reducing unnecessary invasive treatments but also cuts down on medical costs and reallocates resources to more urgent cases.

‍

Loft: How do you set up the tools to continuously improve the product using data and maintain transparency regarding the dataset when submitting it for FDA approval?

‍

Tina: As part of your regulatory submissions, you specify the intended use and indications of use for a product, outlining where it can be applied in clinical settings. If there's a completely new intended scenario or indication, a refiling and new clearance are necessary. However, if the intended use remains unchanged, during the submission of a 510(k) application, you must detail your post-market surveillance methodology—how you plan to monitor the product's performance. As feedback is gathered, if improvements to the algorithm's robustness are made without altering its clinical intended use, re-submission to the regulatory bodies is typically not required. It’s only needed if there is a significant change, such as a complete overhaul of the algorithm's architecture, or if a critical parameter was overlooked in the original model, necessitating a redesign. Therefore, for incremental improvements that adhere to the original intended use and do not fundamentally alter the product, continuous updates can be made without needing to resubmit each time.

‍

Loft: How do you ensure that your AI training datasets continually improve with the introduction of new data over time?

‍

Tina : To ensure that AI training datasets continually improve, it's essential to have a clear data strategy aligned with the specific use case you are addressing. First, define the need or problem the AI model is intended to solve. From there, establish what data is required—this includes specifying the amount of ground truth data, training datasets, and validation datasets needed. Implementing this data strategy involves meticulous data management, such as annotating and cleaning data in a way that enhances the dataset's structure according to your predefined strategy. Randomly adding new datasets without consideration can disrupt this process, so it's crucial to maintain a structured approach. Documenting each aspect of your data strategy is vital. This includes outlining which data points and features are necessary and how they are annotated. When submitting algorithms for regulatory approval, like the FDA's 510(k), you must also include your data strategy and model. This submission should provide transparency about your training dataset, explaining why it's the right choice, detailing the internal verification and final clinical validation datasets, and if applicable, information about clinical study pilot sites. Overall, the goal is to demonstrate a robust, methodologically sound approach to data management that ensures the validity and reliability of your AI model throughout its development and implementation.

‍

Loft: Software as a Medical Device (SaMD) is one of the fastest-growing sectors in the medical device industry. Many startups are looking for ways to incorporate software into imaging (sometimes legacy)  systems, like Microimaging, X-rays, CT imaging, or MR imaging. How do you see the landscape evolve as these technologies are being further developed?  Do you see imaging solutions becoming increasingly platforms for software-driven innovations? Given this context, how do you foresee the evolving landscape of AI in imaging, particularly in terms of monetizing innovation investments?

‍

Tina: Ultimately, it's about defining what constitutes value. There are existing systems, and while the systems themselves are important, the data they produce is equally significant. Nowadays, the positive development is the shift towards a more open ecosystem. This shift is reducing proprietary data formats and communication protocols, and promoting innovation by providing broader access to data through standard interfaces. By adopting a more holistic approach, they can create more impactful and innovative solutions. An example of such innovation is Uber, which transformed the taxi service industry not by focusing on individual drivers but by reimagining the entire user experience. Stepping back to view the whole ecosystem can lead to disruptive innovations that address core user needs more effectively and introduce new business models, significantly enhancing the value delivered by AI and data-driven technologies. Startups and AI experts often possess deep technical knowledge but may lack insight into the larger operational context. Often, algorithm experts may focus narrowly on specific improvements, like enhancing process speed by five minutes. However, it's crucial to step back and consider the overall clinical pathway or workflow. It's not just about a minor time saving but understanding the broader context—identifying pain points, such as why certain errors occur or why particular decisions lead to readmissions. By understanding these broader issues, new AI solutions can be designed to address significant performance bottlenecks, thereby improving overall system efficacy and creating substantial value for the user.

‍

There must be evidence that AI can achieve outcomes that would not be possible with human efforts alone. This doesn't necessarily mean constantly comparing AI with human performance; instead, it's about demonstrating AI's ability to identify issues that are not visible to the naked eye, such as subtle problems in an MRI scan or minor tissue characterizations that a human might miss.

‍

Loft: Within hospital systems, reimbursement strategies often involve multiple stakeholders, each with their distinct priorities. How do we ensure that innovations that save lives, time, and money are implemented, even when the value chain is broken or the incentives for implementation are lacking?

‍

Tina: Yes, it truly is a journey. On the one hand, there are new CPT codes and reimbursements specifically designed for AI integration. To qualify for these, there must be evidence that AI can achieve outcomes that would not be possible with human efforts alone. This doesn't necessarily mean constantly comparing AI with human performance; instead, it's about demonstrating AI's ability to identify issues that are not visible to the naked eye, such as subtle problems in an MRI scan or minor tissue characterizations that a human might miss. These examples illustrate where AI can significantly assist rather than replace, complementing what healthcare professionals can do. Additionally, the rising number of medical cases—driven by a shift towards more preventative care and early screening—increases the workload on healthcare systems. Here, AI can also help by prioritizing cases and preprocessing data to highlight areas of concern, thus enhancing efficiency. These contributions by AI are crucial for gaining reimbursement approval. It's important not just to show that AI can save a few seconds here and there, but to clearly define and prove its value in making significant, measurable improvements to healthcare outcomes. This focus on real value is essential for the continued acceptance and expansion of AI in medical contexts. This awareness is crucial. For instance, in the past, many clinicians didn't recognize the importance of structured data. They often entered information as free-text clinical notes rather than filling out specific fields in the Electronic Medical Records (EMR). For AI, these structured fields are far more valuable because they are readily accessible and reduce the need for natural language processing (NLP), which is prone to errors. When clinicians see the benefits AI can bring and realize that simply filling out a few structured fields can significantly improve AI's performance, they begin to see the value in doing so. This understanding leads to better data practices, enhancing both the AI's effectiveness and the overall healthcare outcomes. Awareness, benefit realization, and the necessary actions to improve data quality go hand in hand.

‍

Loft: It is indeed crucial. You mentioned earlier that data is an important driver in AI innovation. Considering that hospitals generate a vast amount of data daily, it's notable that much of this data is not currently being effectively utilized.

‍

Tina: Exactly. The data exists but isn't always utilized effectively. As you pointed out, the quality may deteriorate if there's a lack of understanding about which data to structure and curate. Instead of adding value, they might just keep accumulating poor-quality data. However, if there's an awareness of the potential value the data holds, there's also a greater consciousness about the quality of data being input to generate that value.

‍

Embracing a culture of rapid experimentation is crucial. Encouraging the mentality of 'experiment fast, learn fast, fail fast' is essential. Trying to achieve breakthroughs with stringent timelines and budget constraints often won't yield the desired outcomes.

‍

Loft: Large organizations often struggle with rapidly evolving innovation and developing groundbreaking, new-to-the-world solutions. How can we build better research and development teams that are capable of achieving breakthrough innovations in AI at the intersection of business, innovation, and research and development?

‍

Tina: Certainly, there are a few key strategies that come to mind when thinking about fostering a culture of innovation. Firstly, embracing a culture of rapid experimentation is crucial. Encouraging the mentality of 'experiment fast, learn fast, fail fast' is essential. Trying to achieve breakthroughs with stringent timelines and budget constraints often won't yield the desired outcomes. For instance, Amazon's approach of setting a clear target for innovation and allowing teams to challenge each other with multiple ideas can be very effective. Secondly, co-creating with customers is vital. Companies like Philips and Siemens Healthineers have learned over time that understanding customer needs cannot be fully achieved from behind a desk. Engaging directly with customers to co-create solutions ensures a deeper understanding of their needs. Thirdly, when seeking feedback for breakthroughs, it’s important not to rely solely on one or two key opinion leaders. Broadening the scope to include feedback from various stakeholders can provide a more comprehensive understanding of common pain points, which may vary significantly across different institutions with different processes. Additionally, having a multidisciplinary team is critical. True diversity and inclusion mean more than just gender diversity; it involves integrating diverse perspectives from various fields such as user experience, product management, marketing, and sales. This diversity in thinking can lead to more innovative solutions. It's also essential not to jump straight to solutions without thoroughly understanding the customer's needs. Observing and asking the right questions to understand what adds value for the customer and what their pain points are can inform more effective solutions. Finally, embracing new technologies such as generative AI can also be a game-changer. Generative AI can act as a sparring partner, helping to generate ideas and facilitate brainstorming, not just for individuals but for teams as well. In summary, innovation requires a strategic approach that combines fast-paced experimentation, customer co-creation, broad feedback, multidisciplinary collaboration, thorough need analysis, and the integration of advanced technologies.

‍

Loft: I've always thought of AI as a co-pilot but considering it as a sparring partner is a really insightful approach.

‍

Tina: I use it, so I like it. In the beginning, it feels like you already know the answers yourself. But over time, as you use AI more, you learn how to use it effectively. You start to realize that not asking the right question can be as detrimental as not asking the right question to a customer. It trains you to think about the questions you're asking based on the answers you receive. Those are some of my thoughts, and I hope they're helpful.

‍

Loft: Thank you for this incredibly interesting interview. Your insights on AI and regulated industries are groundbreaking. I can't wait to continue our conversation in person at some point.

‍

‍