Bitscopic’s Case for Data-First AI

Written by Eugene Geis, Ph.D., Director of Data Science and Research

Data-first AI is a simple premise: our models will never perform better than the data they leverage. In safety-critical domains like healthcare, clinical genomics, and bioinformatics, dirty data is dangerous.

Leaders like Andrew Ng have argued for “data-centric AI,” shifting attention from endlessly tweaking architectures toward improving data quality, labeling, and coverage. In healthcare, this isn’t a competitive hack—it’s a moral and regulatory requirement. Bitscopic’s view is foundational: before you instrument a model, you need rigorous data integrity, provenance, semantics, and interoperability. For our customers, any amplification of noise or “missingness” can mean life and death.

Health records – the raw material for analytics and surveillance – are messy: fragmented EHRs, inconsistent coding, duplicate patients. Reviews continue to find wide variation in completeness and correctness, compromising clinical research, facility monitoring, and AI.

Lies, Damn Lies, and Costly Misaligned AI Models

Financially, Black Book Research estimates that patient-matching errors drive a third of denied claims and cost the average hospital $1.5–2.5 million annually. As medical records were designed around billing, the EHR is a poor substrate for training AI systems. “Garbage in, garbage out” is more than a meme; untrustworthy clinical data becomes a patient-safety liability.

When institutions ignore data quality and put modeling first, failures become public. The most famous example is IBM Watson for Oncology. Internal documents revealed “unsafe and incorrect treatment recommendations,” and critics noted the system was trained on hypothetical cases and expert preferences rather than broad real-world patient data.

At MD Anderson Cancer Center, a collaboration to build the Oncology Expert Advisor ended after an audit found $62 million spent with no progression to clinical trust. The reputational damage has become a cautionary tale: sophisticated ML cannot rescue a system built on narrow, poorly governed data.

Financial Waste: Messy EHR data (fragmented, inconsistent, mismatched records) costs the U.S. healthcare system billions annually through denied claims and duplicated efforts.

Safety & Liability: In safety-critical domains (healthcare, genomics), dirty data is dangerous and can lead to medical malpractice (e.g., amplification of noise/untrustworthy signal).

Many label such failures “bias,” but they more often reflect a lack of inferential rigor across measurement, pipelines, transformations, and outcome evaluation. When the Bitscopic R&D team reviewed clinical ML papers from 2020–2023, fewer than 5% of more than 70 classification studies discussed safeguards for Type I and Type II errors. It is no surprise MIT’s Innovation magazine concluded: “Hundreds of AI tools have been built to catch covid. None of them helped.“

Even large vendors have stumbled. The Epic Sepsis Model (ESM), deployed in hundreds of hospitals, performed poorly in an external validation study in JAMA Internal Medicine. University of Michigan researchers found it missed two-thirds of sepsis cases while generating high false-alert volumes. Summaries noted it “performed much worse than what the vendor tells its customers,” correctly ranking patient risk only about 63% of the time. The problem wasn’t just algorithmic weakness—it was the absence of transparent, data-grounded validation. The result: wasted clinician time and eroded trust in AI-enabled decision support.

Move Fast and Break Your Credibility

Bias also emerges from data choices that hard-code inequities. A widely used care-management algorithm from Optum underestimated the needs of Black patients because it used healthcare cost as a proxy for illness severity; historically lower spending led the model to conclude these patients were “healthier,” reducing resource allocation. Regulators opened probes. “Move fast and break stuff” may work in consumer tech, but in clinical informatics, relying on convenient proxies instead of designing datasets that reflect clinical reality and health equity creates real harm.

Across domains, the financial stakes of bad data are enormous. A widely cited analysis estimates a $3.1 trillion annual drag on the U.S. economy from rework and inefficiencies. In healthcare, patient-identity and record-quality issues drive billions in denied claims and duplicated tests. Once AI is layered on top, every data defect compounds: mis-triaged sepsis, biased risk scores, unnecessary denials of care. Any speed gained by rushing a model evaporates under re-implementation costs, regulatory attention, and reputational damage.

AI Cannot Outperform Data Quality

Regulators are increasingly explicit that trustworthy AI begins with data validity and provenance. The FDA’s discussion papers on AI/ML-based medical software emphasize robust real-world data and ongoing monitoring. NIST’s AI Risk Management Framework highlights data quality, validity, and reliability as core characteristics of trustworthy systems. Speed at the expense of data rigor is a roadmap to scrutiny when models influence diagnoses, coverage decisions, or public-health responses.

Bitscopic’s thesis is that sustainable algorithmic stewardship requires data-first AI. That means investing in the unglamorous work of harmonizing real-world clinical, public-health, and genomic data into a clean, semantically rich, interoperable foundation. Models are not trained until we verify (1) robust provenance pipelines, (2) standardized vocabularies and ontologies, (3) cross-facility patient identity and cohort resolution, and (4) continuous data-quality monitoring that surfaces drift and anomalies before they contaminate inference. These principles shape our work in genomics (PraediGene), facility-level surveillance (PraediAlert), and national biosurveillance (PraediCo): models are only the last mile of a much larger data and semantics marathon.

This is the first installment of a series on what data-first AI means for genomics, infection prevention, and national public-health intelligence. We’ll explore how to design datasets and semantic layers that withstand regulatory scrutiny, how to measure and monitor data quality in the wild, and how to define requirements for AI systems clinicians and public-health officials can trust. If your organization wants to move beyond “shiny demo models” toward reliable, high-impact AI, the first step isn’t a bigger model—it’s a stronger data foundation. That’s the foundation Bitscopic has committed to reinforcing.

Download the complete Data-First AI brief now for free

With AI in healthcare, model performance is capped by the integrity of the data underneath it. Download the full article by Eugene Geis, Ph.D., which goes deeper on real-world failure modes, validation failures that erode clinical trust, and Bitscopic’s data-first requirements for trustworthy AI across healthcare, clinical genomics, and public health.

This is a free download. No form. No email required.