Atul Gawande

Diagnostic uncertainty remains medicine's ground state, yet most AI clinical decision support tools are built to project confidence rather than calibrate it. They give answers when what clinicians actually need is help knowing when they don't know enough. Current AI diagnostic tools optimize for accuracy on clear cases but fail precisely where medicine fails—in ambiguous presentations like the young woman with the red leg who might have cellulitis or might be dying of necrotizing fasciitis.

An uncertainty-forward diagnostic companion that doesn't just suggest diagnoses but explicitly maps the decision landscape—showing what findings would shift probability dramatically, what the cost of being wrong is in each direction, and when the honest answer is 'I don't know, but here's what would tell us.' It would be trained not on resolved cases but on cases that were initially misdiagnosed, learning the specific patterns where confident clinicians get fooled.

Medical errors still kill between 250,000 and 400,000 Americans annually, yet error reporting systems remain clunky, punitive, and disconnected from learning. Doctors don't report because the systems feel like documentation of failure rather than architecture for improvement. Reporting tools from 2003 remain in use, designed around compliance rather than understanding.

A frictionless, voice-enabled error and near-miss capture system integrated into the clinical workflow—something a resident could speak into for thirty seconds after a fumbled central line, that uses AI to identify patterns across institutions, that feeds back anonymized learning within days rather than years. It would generate case studies automatically, identify systemic vulnerabilities, and connect clinicians who've made similar mistakes to learn from each other.

The autopsy rate has collapsed to under 5% while diagnostic error rates remain stubbornly around 40% in cases that do get autopsied. We've lost our primary quality control mechanism for diagnosis precisely when we most need to understand where our increasingly complex medicine goes wrong.

A network of regional rapid-feedback autopsy centers with streamlined family consent processes, AI-assisted pathological analysis, and a commitment to returning diagnostic findings to treating physicians within 72 hours with educational context. Pair this with a cultural intervention—training modules that help physicians approach autopsy requests not as admissions of failure but as the final act of care. Financially, support it through malpractice insurers who have every incentive to understand error patterns.

End-of-life decision-making remains chaotic, with families and patients making choices in crisis without understanding what they're actually choosing between. Patients and families hear options without understanding what they actually mean, what they will feel like to live through, or what outcomes they're really accepting.

A decision architecture platform for serious illness—not advance directives that no one reads, but interactive tools that walk patients through realistic scenarios before crisis hits. Use narrative and video to show what ICU dying actually looks like versus hospice, what ventilator life means versus comfort care. Help patients articulate values around suffering, dependency, and being a burden in ways that translate to specific clinical decisions.

Chronic pain affects over 50 million Americans, yet treatment remains primitive—oscillating between dangerous opioids and dismissive suggestions to 'learn to live with it.' Current digital pain management tools focus on tracking symptoms rather than the complex psychological-physical-social matrix that actually produces suffering.

A comprehensive pain reconceptualization platform that treats chronic pain as the multidimensional phenomenon it actually is—part nerve signal, part learned behavior, part attention, part depression, part social isolation. Combine biofeedback, cognitive behavioral components, movement coaching, social connection features, and careful medication guidance in one integrated system. Design it with the understanding that patients dismissed by medicine need validation before they can engage in treatment.

Five billion people lack access to safe surgical care, concentrated in low-income countries where training infrastructure is weakest. Mission trips provide episodic help but not sustainable capacity. The surgical techniques learned through apprenticeship don't scale across global health disparities.

A surgical training amplification system combining high-fidelity simulation, remote mentorship infrastructure, and AI-assisted procedural guidance. A trainee in rural Rwanda could practice a hernia repair on sophisticated simulation, receive real-time guidance during actual surgery from an experienced surgeon watching through smart glasses, and get structured feedback after each case. Let one experienced surgeon's expertise reach dozens of trainees across distances.

Healthcare workers are burning out at unprecedented rates—over 60% reporting symptoms—and the interventions remain superficial: meditation apps, resilience training, yoga classes that blame clinicians for failing to cope with impossible systems. The documentation burden alone consumes hours that should go to patient care.

A practice restructuring consultancy that redesigns workflows rather than another wellness app. Drawing from industrial engineering and surgical checklist lessons, enter health systems and identify the specific sources of cognitive overload, interruption, documentation burden, and moral distress—then systematically eliminate them. This means ambient AI documentation that actually works, care team restructuring that matches task to appropriate training level, and protected time for human connection.

Obesity treatment remains trapped between ineffective willpower-based interventions and bariatric surgery that's invasive, expensive, and available to only a fraction of those who need it. We have no system for helping people know which intervention matches their specific physiology, and obesity is being treated as one disease when it's dozens of conditions.

A precision obesity care platform that starts with comprehensive metabolic and behavioral phenotyping—not just BMI, but appetite hormones, eating patterns, psychological relationships with food, social and environmental factors—then matches individuals to the intervention most likely to work for them. For some that's medication, for some surgery, for some intensive behavioral support, for most some combination.

Medical education still relies on the uncomfortable truth that learning requires practice on real patients. Simulation technology remains peripheral to training, used for certification checkboxes rather than genuine skill development before patient contact.

A mastery-based procedural training system where residents cannot attempt procedures on patients until they've demonstrated consistent competence in simulation—not passing a test once, but proving they can reliably perform under varying conditions, with complications, with distractions. AI assesses technique in real-time and provides immediate feedback. Include not just manual skills but cognitive aspects: knowing when to call for help, recognizing you're out of your depth, communicating with patients. Track long-term outcomes to continuously improve what simulation competence predicts.

The conversation about illness remains medicine's most undertaught skill. Doctors communicate certainty when they should express uncertainty, options when they should offer guidance, information when they should share feeling.

A systematic communication curriculum embedded throughout medical training, using AI-assisted role-play that provides immediate feedback on language, tone, pacing, and emotional attunement. Practice with simulated patients who respond to how they're being spoken to. Capture what the best communicators actually do differently—the pause before difficult news, the question that checks understanding, the acknowledgment that something is terrifying—and help others learn those specific moves.