Updates

A randomised trial of 249 physicians in Indonesia, Kenya and the Netherlands (npj Digital Medicine) found GPT-4o access raised vignette scores by 7 to 18 percentage points. The control doctors were not allowed guidelines or internet searches.

Why it matters: The result shows GPT-4o beats unaided memory, not that it beats a doctor with UpToDate or a guideline open.

In Academic Radiology, five radiology residents read 1,861 chest radiographs with four commercial AI tools. Accuracy did not improve overall and fell for pleural effusion and nodules because of false positives, though three of five residents read 6 to 17 seconds faster.

Why it matters: Readers felt more confident while accuracy stayed the same, which is the pattern that leads to automation bias.

NMC chairperson Dr Abhijat Sheth said close to 50,000 doctors finished structured online AI training this year. He repeated that responsibility for clinical judgement stays with the doctor.

Why it matters: Expect AI literacy to become a normal part of training and CME, and you stay accountable for any AI output you use.

The FDA's device centre released 'Considerations for the Regulation of Generative AI-Enabled Medical Devices', covering a two-axis risk model, premarket benchmarking plus clinical confirmation, and postmarket monitoring. It is a discussion paper, not a rule, and comments close on 19 October 2026.

Why it matters: This is the first sign of how LLM-based clinical tools may be cleared in the US, and Indian developers selling there will follow it.

The report 'AI in MedTech' was released by Union Health Minister J.P. Nadda at India Medical Device 2026. It lists data readiness, regulation and reimbursement as the three gaps between pilots and wide use.

Why it matters: Diagnostic AI is most likely to reach district and secondary hospitals first, so PG trainees there will meet it early.

The guidance explains how software, including AI/ML, cloud and mobile apps, is handled under the Medical Devices Rules 2017. AI products must show performance in at-risk populations and care settings and address bias and Indian demographic representation, and CDSCO describes it as clarifying existing rules.

Why it matters: Clinical AI sold in India now has a written documentation standard, which gives hospitals a checklist when buying tools.

HOPPR's fifth foundation model reads 3D chest CT volumes and describes findings across lung, upper abdomen, heart, mediastinum and soft tissue, including pneumothorax and pulmonary embolism. It is offered through the company's deployment services to radiology departments.

Why it matters: CT report drafting from whole volumes is moving from research into vendor products.

The FDA chose Dexcom as the first participant in TEMPO, a pilot with CMS to gather real-world evidence on digital health devices, including AI insights, for chronic disease care. It plans up to ten more participants in each of four clinical areas.

Why it matters: The US is testing a route where devices are monitored on real-world outcomes instead of only premarket data.

A two-day review (17 to 18 July) of Ayushman Bharat PM-JAY and ABDM set AI, digital services and claims financing as priorities. It also covered linking medical college hospital information systems to ABDM standards and an NMC dashboard for college assessment.

Why it matters: Hospital records and claims in your institution are likely to be pushed toward ABDM-linked digital workflows.

A cluster-randomised trial of an LLM tool ('AI Consult') inside the record system, in 16 primary care clinics with over 9,600 patients, improved documentation and treatment planning. Treatment failure was 2.2% with the tool and 2.0% without, and no harm signal was seen.

Why it matters: One of the few real-world trials in a low-resource setting similar to Indian primary care, showing better process but no proven outcome gain yet.

A cluster-randomised trial at two UC San Diego emergency departments used an LLM to check SEP-1 sepsis care and give feedback. Compliance was 82.9% versus 70.1%, with no significant difference in mortality or ICU admission.

Why it matters: The gain was in meeting the bundle and documenting it, not in survival.

The radiology model drafts narrative reports using the clinical question, patient history and prior studies, and was trained on about 6 million imaging studies. It is not cleared for clinical use and is available for research access only.

Why it matters: Prior-comparison report drafting is the feature to watch, but this one is not yet a clinical product.

Seven AI devices were tested between April 2025 and March 2026. Findings include the need for post-market monitoring, watching how human oversight changes over time, and tying performance thresholds to clinical importance. Phase 3 is funded at 1.2 million pounds a year for 2026 to 2029.

Why it matters: The sandbox findings spell out how to judge a clinical AI tool beyond its headline accuracy.

A provisional agreement on 6 May, confirmed by the Council on 13 May, moves AI embedded in regulated products such as medical devices to 2 August 2028. Stand-alone high-risk systems move to 2 December 2027.

Why it matters: CE-marked AI tools you may see in Indian hospitals get about a year more before AI Act obligations apply on top of MDR.

Across five US academic centres, ambient scribes saved about 16 minutes of documentation per 8 hours of patient care, and 27.3 minutes for clinicians who used them on at least half of visits. A companion JAMA Network Open study of 263 clinicians found burnout fell from 51.9% to 38.8% after 30 days.

Why it matters: Gains are modest on average and depend on how consistently the tool is used.

Dr Abhijat Sheth said AI is already part of the clinical environment and doctors should understand its limits while keeping clinical judgement. He said training only in the traditional framework risks a gap between what is taught and what is practised.

Why it matters: This is the regulator's public stance: learn AI, but accountability stays with you.

Launched by Medical Dialogues with the National Medical Forum, with NBEMS, ICMR and AMTZ among the partners. It aims to teach doctors to read AI-assisted diagnostics, interpret algorithm alerts and cut documentation work.

Why it matters: A doctor-led AI literacy programme aimed at Indian clinicians, though it is not an official regulator requirement.

The agreement brings IndiaAI's compute and dataset platforms together with ICMR's research expertise, using ICMR's MIDAS framework for responsible AI in health.

Why it matters: It is meant to widen access to Indian health datasets and compute for researchers, which may help thesis and research projects.

The product helps with documentation, writing and medical research, and comes with a benchmark called HealthBench Professional. OpenAI reports that 99.6% of nearly 7,000 real-world clinical conversations were rated accurate and safe; that is the company's own figure.

Why it matters: It is limited to verified US clinicians for now, and the accuracy figure has not been independently confirmed.

In a three-arm randomised trial of 2,345 people, correct naming of a skin condition was 7.9% with web search, 22.8% with an AI tool and 36.2% with dermatologist-curated differentials. Understanding of next steps did not improve.

Why it matters: Patients arriving with an AI-suggested diagnosis are likely to be more confident but still often wrong.

In a 100-patient feasibility study at Beth Israel Deaconess, patients text-chatted with AMIE before appointments under live physician supervision. No safety stops were needed, the final diagnosis was in AMIE's top 7 in 90% of cases, and it was the single top diagnosis in 56%.

Why it matters: It is early, text-only and uncontrolled, and physicians still did better on practicality and cost.

SAHI sets governance and validation direction for health AI, and BODH, built by IIT Kanpur with the National Health Authority under ABDM, lets models be tested on real-world data while protecting privacy. Union Health Minister J.P. Nadda launched both at the India AI Impact Summit.

Why it matters: This is India's first national framework for how health AI should be validated and governed.

The five-day summit (16 to 20 February) at Bharat Mandapam had more than 840 exhibitors, and the Prime Minister named healthcare among sectors AI is changing.

Why it matters: The summit is where the national health AI strategy was announced the next day.

In the Swedish randomised trial of over 100,000 women, AI-supported screening cut interval cancer diagnoses by 12% with similar false positives (1.5% vs 1.4%). It also found 16% fewer invasive and 27% fewer aggressive-subtype cancers between screens.

Why it matters: This is the strongest randomised evidence so far that AI in screening can change cancer outcomes, not only workload.

MedGemma 1.5 4B reads CT, MRI and histopathology images and runs locally, and MedASR is a medical dictation speech model that Google says makes 58% fewer errors on chest X-ray dictation than general models.

Why it matters: Open models that run on a hospital's own hardware may suit privacy rules better than cloud tools.

The offering adds connectors for databases such as ICD-10 and PubMed and, for US users, a HealthEx link to patient records. Anthropic says health data is not used for model training.

Why it matters: It follows OpenAI's health launch, and the record-linking features are US-only for now.

In Open Forum Infectious Diseases, qXR reading of 2,745 X-rays in a tribal population flagged 363 presumptive cases, 162 were confirmed, and notifications rose 80.21% against baseline. It also found 51 asymptomatic people who symptom screening would have missed.

Why it matters: It is an Indian public-health result for AI-read X-rays in active TB case finding, though not a randomised trial.

The FDA now allows enforcement discretion for tools giving a single output when only one recommendation is clinically appropriate. Software that analyses medical images to give diagnostic recommendations remains regulated as a device.

Why it matters: Many text-based clinical support tools can avoid FDA review, while image-reading AI cannot.

The programme has 20 live online modules with faculty from Mayo Clinic, Harvard, Oxford, IISc and IIM Lucknow, and needs 75% live attendance for a certificate. Eligible people include current NBEMS trainees, alumni from 2020 and accredited-department faculty.

Why it matters: This is a free, certified AI course aimed at DNB trainees.

UpDoc, a type 2 diabetes insulin dosing support tool, was cleared through 510(k) K253281 with a predetermined change control plan. The LLM handles conversation while fixed deterministic logic calculates the insulin dose.

Why it matters: It shows how LLMs are being cleared for now: the language model talks, and rule-based code makes the dosing decision.

Consent manager rules apply from 13 November 2026, and most duties for data fiduciaries apply 18 months after notification, on 13 May 2027. The rules require reasonable security safeguards and set no separate health-specific category.

Why it matters: Hospitals and clinics are data fiduciaries, so patient data used with AI tools will need consent and security controls.