> ## Content Index > Fetch the complete content index at: https://www.healthdatacon.io/llms.txt > Use this file to discover other available public pages before exploring further. # Fatema Ravatwala and Julia Rehman Put Readiness Before Healthcare AI Adoption - URL: https://www.healthdatacon.io/fatema-ravatwala-and-julia-rehman-put-readiness-before-healthcare-ai-adoption/ - Published: 2026-10-16T07:16:00.000Z - Updated: 2026-10-16T07:15:59.000Z - Description: At HealthDataCon 2026, Fatema Ravatwala and Julia Rehman outlined how data quality, named accountability and local validation can help health systems establish readiness before adopting clinical AI. - Author: Health Data Management - Tags: EMEA Before a health system switches on an AI tool, who can show that it is safe for its patients, equitable across its population and accountable when something goes wrong? At HealthDataCon 2026, primary presenter Fatema Ravatwala and co-presenter Julia Rehman made that question the starting point for adoption. Their EMEA session, “Ready Before AI—Data, Accountability, and Population Fit in Regulated Health Systems,” presented a practical framework supported by published research. Ravatwala, MBA, BDS, eFACHDM, is a healthcare strategy and digital health consultant whose clinical background is dentistry. In an earlier [interview](https://www.womenentrepreneursreview.com/leader/fatema-ravat-an-experienced-dentist-managing-business-operations-to-connect-patients-to-specialized-dentists-vid-576.html?ref=healthdatacon.io), she described practicing in India and Dubai before pursuing healthcare management. Rehman, DHA, FACHE, eFACHDM, is founder and chief operating officer of Kota Kompany LLC. Ravatwala developed the accountability and population-fit arguments; Rehman addressed organizational readiness and implementation. They distinguished their framework from the researchers’ evidence and made a clear proposition: responsible adoption starts with an organization’s capacity to use, question and govern a tool. ![Julia Rehman co-presenting at HealthDataCon 2026.](https://storage.ghost.io/c/91/db/91db0927-dd9a-4578-9944-921da70f6885/content/images/2026/10/julia-rehman-healthdatacon-2026.jpeg) Julia Rehman at HealthDataCon 2026\. Photo credit: Julia Rehman / HDM’s HealthDataCon 2026. ## Make the weakest dimension visible Rehman described three recurring mismatches: technology advances faster than governance, investment exceeds workforce capability, and expectations outrun data quality. A purchased license does little to resolve an absent escalation pathway or records that omit information a model needs. The assessment separates six dimensions: governance and leadership; infrastructure and connectivity; data quality and standards; workforce capability; interoperability; and culture and change readiness. Each requires evidence. A committee without stop authority, an unreliable connection or one indispensable analyst can expose a different weakness. The speakers recommended making the weakest dimension visible to leadership instead of hiding it inside an average. This was their practical assessment approach, not a claim that the illustrative scorecard had been independently validated. Data quality received particular attention. Complete, accurate, consistent, timely, standardized and traceable information enables scrutiny of both inputs and outputs. “AI does not clean your data. It amplifies it,” Rehman said. A [2019 study](https://pubmed.ncbi.nlm.nih.gov/31649194/?ref=healthdatacon.io) showed why even an apparently objective target can mislead. A population-health algorithm used healthcare spending as a proxy for need. Because less was spent on Black patients with comparable illness, they were disadvantaged. Researchers estimated that correcting the disparity would increase Black patients’ share among those selected for extra help from 17.7% to 46.5%. The governance question extends beyond accuracy: what was the model taught to predict? ![Four AI accountability roles: executive sponsor, clinical owner, data steward and model owner.](https://storage.ghost.io/c/91/db/91db0927-dd9a-4578-9944-921da70f6885/content/images/2026/10/ai-accountability-roles.png) Four accountability roles. Slide credit: Fatema Ravatwala and Julia Rehman / HealthDataCon 2026. ## Put names and authority behind accountability Ravatwala translated accountability into four roles. The executive sponsor sets risk tolerance, answers to the board and funds continued operation. The clinical owner evaluates clinical validity and can override the output. The data steward owns data quality, lineage and review of the prediction target. The model owner monitors performance, investigates discrepancies and manages escalation or retirement. “Accountability has to be assigned before go live,” she said. Her emphasis was on identifiable people with usable authority. An override exists in practice only when clinicians can exercise it without fear of blame for questioning an alert. Small organizations may combine some responsibilities, Ravatwala acknowledged. She recommended separating clinical and model ownership so one can challenge the other. That is a proposed organizational safeguard, rather than a universal staffing requirement. Abu Dhabi’s [AI standard](https://www.doh.gov.ae/-/media/Feature/Resources/Standards/2025/Responsible-AI-Standard-V1.ashx?ref=healthdatacon.io) provides a regulatory foundation. It calls for defined responsibilities, designated owners, documented decisions and clinician override mechanisms. Its wording permits roles assigned to individuals or teams; the speakers’ four-person map makes their own accountability approach more specific. The standard also addresses representative data and continuing oversight. Ownership must continue after deployment. Someone needs time, tools and authority to examine unexpected results, investigate harm and pause a system. A governance document cannot perform those duties by itself. ## Read sepsis evidence with its limits Two sepsis studies illustrated why leaders must examine both prediction and response. They studied different products using different methods, so they cannot establish that one governance approach caused better outcomes. A Michigan [external validation](https://jamanetwork.com/journals/jamainternalmedicine/fullarticle/2781307?ref=healthdatacon.io) examined 38,455 hospitalizations. The model’s area under the receiver operating characteristic curve was 0.63\. At the studied threshold, sensitivity was 33% and positive predictive value was 12%: about two-thirds of sepsis cases were missed, while roughly one in eight flagged hospitalizations involved sepsis. The threshold would have flagged 18% of hospitalizations. These were retrospectively evaluated, hypothetical alerts, not alerts delivered to clinicians during the study. The findings demonstrate a performance and potential workload problem at that institution; they do not establish patient harm from responding to those alerts. A separate [prospective study](https://www.nature.com/articles/s41591-022-01894-0?ref=healthdatacon.io) monitored 590,736 patients across five hospitals. Its outcomes analysis focused on 6,877 patients with sepsis identified before antibiotics. Confirmation within three hours was associated with an adjusted mortality reduction of 3.3 percentage points, or 18.7% relatively, compared with later or absent confirmation. Rehman emphasized that this was an observational association, not proof of causation, and noted author financial ties to the developer. Her operational lesson was that a prediction requires a response process: who receives it, when they review it and what they can do next. Organizations must evaluate that process alongside model performance. ## Test whether performance travels Ravatwala challenged the assumption that validation elsewhere establishes suitability locally. Population characteristics, available data, equipment and clinical routines can all alter performance or usefulness. A South African [tuberculosis study](https://pmc.ncbi.nlm.nih.gov/articles/PMC11339183/?ref=healthdatacon.io) compared 12 commercial chest-radiograph AI products using 774 participants from a national prevalence survey, including 258 with confirmed tuberculosis. Thresholds differed across products and versions, and performance was generally poorer in older people, people living with HIV and those with previous tuberculosis. This was a case-control evaluation, not a demonstration that every product would fail in clinical deployment. It supports checking the actual version and the people a service expects to treat. Rehman also presented a constructive example: Sweden’s randomized MASAI mammography trial. The [2025 analysis](https://pubmed.ncbi.nlm.nih.gov/39904652/?ref=healthdatacon.io) reported 338 cancers detected among 53,043 participants receiving AI-supported screening, compared with 262 among 52,872 receiving standard double reading—6.4 versus 5.0 per 1,000. The evidence has continued to develop. A [2026 analysis](https://pubmed.ncbi.nlm.nih.gov/41620232/?ref=healthdatacon.io) met the trial’s non-inferiority criterion for cancers diagnosed between screening rounds: 1.55 versus 1.76 per 1,000 participants. The confidence interval did not establish a statistically significant reduction in that rate. The trial supports careful evaluation within a functioning screening program; it does not establish a mortality benefit or automatic transferability to another population. ![Five perspectives for AI adoption: executive, clinical, nursing, data and patients.](https://storage.ghost.io/c/91/db/91db0927-dd9a-4578-9944-921da70f6885/content/images/2026/10/ai-adoption-five-lenses.png) Five perspectives on adoption. Slide credit: Fatema Ravatwala and Julia Rehman / HealthDataCon 2026. ## Turn readiness into a 90-day work plan The closing exercise made the framework concrete. A hypothetical 400-bed hospital receives a free first-year sepsis model trained elsewhere. Its electronic record covers only the emergency department, laboratory results arrive on paper and nursing observations are inconsistent. The board wants immediate deployment. Rehman’s answer was “Not yet,” accompanied by a route forward. Over successive quarters, the hospital would name owners and repair data foundations, test retrospectively on local records, run the model silently without acting on alerts, and make an evidence-based deployment decision with response procedures and exit terms agreed. The first 90 days begin that work; they are not a promise to launch. Convene clinical, health information management and IT leaders. Assign ownership. Assess the six dimensions and disclose the weakest. Establish a data-quality baseline, narrow the scope to available capability, define pilot success before results arrive, and prepare monitoring and escalation before go-live. The speakers also asked leaders to examine adoption through five perspectives: executives, clinical leaders, nursing and frontline staff, information and data professionals, and patients and communities. Procurement questions should cover the clinical problem, training population, local validation, data availability, workflow, monitoring, named accountability, ongoing cost and an exit plan. Asked which single improvement deserved funding first, Rehman prioritized accessible, reliable data. Ravatwala agreed and emphasized validation in the intended population. Those priorities give leaders something tangible to commission before purchasing another capability: evidence about their own readiness, their own patients and the people who will remain responsible.