Things & Thinks-Issue LXVII
📚Research Digest
Microsoft's MAI-DxO: Orchestrating Smarter, Cheaper Clinical Reasoning
What it is about
This study presents a new benchmark by Microsoft—Sequential Diagnosis Benchmark (SDBench)—designed to evaluate AI systems in realistic clinical diagnostic scenarios. Unlike traditional multiple-choice evaluations, SDBench simulates real-world medical encounters by requiring stepwise reasoning: starting from a brief case summary, the AI (or physician) must iteratively ask questions and request tests before making a diagnosis. It uses 304 challenging NEJM clinicopathological conference cases and includes cost estimation for each test or physician interaction. The authors also introduce MAI Diagnostic Orchestrator (MAI-DxO), an AI framework inspired by team-based clinical reasoning. MAI-DxO simulates a virtual panel of physician roles, enabling strategic, cost-conscious diagnostic decisions. When paired with OpenAI’s o3 model, MAI-DxO achieved up to 85.5% diagnostic accuracy—far exceeding the 20% average of experienced physicians—while also reducing costs by up to 70% compared to non-orchestrated AI models.
What it means
By demonstrating that structured orchestration boosts both diagnostic accuracy and cost-efficiency, the authors suggest that AI systems can eventually serve as powerful diagnostic collaborators, especially in under-resourced settings. The model-agnostic nature of MAI-DxO provides resilience to rapidly evolving AI tools, avoiding constant retraining. Moreover, tools like MAI-DxO could assist not only in improving care delivery but also in medical education, patient triage, and global health access—particularly where experienced specialists are scarce. Future efforts will need to test such systems in routine, everyday clinical environments and integrate modalities like imaging for broader diagnostic coverage. Ultimately, this signals a shift toward AI-augmented, team-based medical diagnostics that blend efficiency, accuracy, and accessibility.
SensorLM
What it is about
This next one, by Google researchers, introduces SensorLM, a family of foundational models that align wearable sensor data with natural language. Recognizing the challenge of interpreting real-world sensor streams due to the scarcity of paired sensor-text data, the authors developed a hierarchical captioning pipeline to systematically generate textual descriptions capturing statistical, structural, and semantic aspects of raw sensor data. This enabled the creation of the largest known sensor-language dataset, covering over 59.7 million hours of wearable recordings from 103,000+ individuals. SensorLM extends well-known multimodal architectures like CLIP and CoCa into a unified sensor-language framework. Through rigorous evaluations in domains like human activity recognition and healthcare, SensorLM outperforms current methods across zero-shot, few-shot, and cross-modal tasks, demonstrating strong generalization, scaling properties, and efficient learning with limited labels.
What it means
SensorLM;s zero-shot and few-shot abilities could drastically reduce the need for labeled data, accelerating deployment in fields like personalized health monitoring, elderly care, and activity tracking. The hierarchical captioning technique could influence future dataset construction in other sensing modalities. Moreover, by integrating sensor data into the same multimodal space as text and vision, SensorLM paves the way for more intuitive human-AI interaction, where users can query sensor states or trends in plain language. This research signals a move toward foundation models for ambient intelligence, making real-world sensing more explainable, scalable, and adaptable.
🖇Digital Healthcare News
#GenAI and #BigTech in #Healthcare
Elon Musk’s xAI launched ‘Grok for Government’ for healthcare and science use.
Physician networking company Doximity released a free AI scribe, highlighting the subsector's lack of technical moat.
Regulatory Brief
UK NHS will develop AI technology to scan NHS systems to flag safety issues in real time and trigger crucial inspections earlier, as part of the UK government’s Plan for Change to shift NHS services from analogue to digital under the 10 Year Health Plan.
Aktiia received FDA 510(k) clearance for its over-the-counter cuffless blood pressure monitor, G0 Blood Pressure Monitoring System, also known as the Hilo Band.
Neu Health, a smartphone platform for Parkinson's disease and dementia, announced it received FDA 510(k) clearance for its smartphone-based tremor measurement module, which aims to quantify tremor in adults with mild to moderate Parkinson's disease.
Recommended by LinkedIn
Pharma/Device Brief
Revolution Medicines collaborated with Iambic Therapeutics’ AI drug discovery platform, striking a multi-year technology and research collaboration.
Funding, Deals, Mergers & acquisitions
Ambience Healthcare, an ambient AI documentation company, raised $243M
OpenEvidence, an AI-enabled medical research aggregate platform for doctors, raised $210M
AI-enabled imaging company Aidoc raised $150M, partly to accelerate the development of CARE, a clinical-grade foundation model.
Slingshot AI raised $53M to launch an AI chatbot trained on sessions with human therapists
Mandolin, a platform that uses AI automation to enhance access to specialty drugs, raised $40M
Arbital Health, an AI platform to help manage value-based care contracts, raised $31M
Charta Health, an AI-enabled platform that automates billing and coding workflows, raised $22M
Consumer Digital Health & Other News
Samsung acquired digital health company Xealth, a platform that helps providers manage digital tools,
📙Longread of the Month
Not really a long read but definitely worth a thought
🦜Tweet of the Month
This is a funny one...
📊Chart of the Month
According to this McKinsey review, pharma leads in AI economic potential with high R&D acceleration opportunities (40-120% gains)