Mustafa Suleyman’s 2016 Vision for Technology in Patient Care

At the King’s Fund Digital Health and Care Congress in July 2016, DeepMind co-founder Mustafa Suleyman discusses how technology might support clinicians and patient care. His examples and plans reflect that historical moment.

Suleyman describes DeepMind’s general-purpose AI research and its aim to learn useful behavior from data. He situates the healthcare work within a research organization that also wants real-world applications.

He introduces reinforcement learning through an agent that acts in an environment, receives observations and a reward, and adjusts its behavior. This is a research framework, not a description of how every healthcare tool works.

A game example illustrates learning through repeated interaction. The system receives a visual input and available actions, then improves its play through experience rather than being given a complete hand-written strategy.

He then discusses AlphaGo and the difficulty of manually specifying all useful knowledge for a complex game. The point is that learned representations can support tasks with large search spaces.

Turning to health, he says the organization wants technology to have a beneficial impact while developing a sustainable operating model. He describes healthcare as a setting with urgent information and coordination problems.

He argues that clinically relevant information can be fragmented across paper, charts and systems, making it hard for staff to obtain a timely picture. He compares this with consumer software designed for quick use.

Suleyman acknowledges a history of disappointing healthcare technology projects. He argues that the development process must begin with users and real clinical workflows, rather than technology in isolation.

He describes short cycles of observing work, developing prototypes, testing them with users and learning from feedback. Clinicians should identify problems that matter in their day-to-day work.

One target is delayed recognition and communication of patient deterioration. The proposed software aims to make information available to the people who need to assess and act on it.

He discusses acute kidney injury and an existing alert algorithm used in the NHS. The project he describes aims to organize relevant test results and support clinical communication; an alert is not a diagnosis or treatment plan.

A proposed mobile interface lists patients with alerts and lets staff examine information such as blood-test results. These are product demonstrations and goals from 2016, not evidence of current clinical performance.

He describes a possible future view of patients by ward or specialty, with a timeline and relevant clinical information. The intended benefit is a shared view for the people involved in care.

Communication is the second half of the problem: identifying a risk only helps if the right clinician can review it and coordinate an appropriate response. The presentation describes a workflow, not an autonomous medical decision.

Suleyman also introduces research on eye scans and conditions including diabetic retinopathy and age-related macular degeneration. He presents earlier detection as a potential research goal, not a proven benefit for every patient.

He discusses work with clinical collaborators using optical coherence tomography scans. The aspiration is to help specialists handle information and identify conditions that need attention, subject to validation.

The speaker outlines principles for the program: patient involvement, clinician leadership, interoperability and external scrutiny. He invites feedback from people who would use or be affected by the tools.

He argues that patients should have a stronger role in their own care and accountability. He also calls for nurses and doctors to participate in testing and prioritizing projects.

On interoperability, he advocates open standards and a common way for applications to access appropriate information across different hospital systems. Such integration also requires security, governance and legal safeguards.

He describes independent reviewers who can scrutinize the program and speak publicly. The stated commitment to scrutiny should be assessed against subsequent implementation and independent findings.

Later context: in 2017, the UK Information Commissioner concluded that the Royal Free trust’s provision of patient records to DeepMind for Streams clinical safety testing did not fully comply with data-protection law. The finding concerned that specific arrangement, not all healthcare AI projects. See the [Information Commissioner’s 2017 investigation](https://ico.org.uk/media/action-weve-taken/undertakings/2014353/undertaking-cover-letter-revised-04072017-to-first-person.pdf).