A Rectal Cancer Patient, Three Conflicting Truths

Picture a single patient with rectal cancer. On the staging MRI, the tumor is described as 5 cm from the anal verge. The colonoscopy report measures it at 3 cm. The operative note reads "approximately 4 cm from the dentate line." All three statements describe the same anatomical fact. None of them agree.

This is not a documentation error. It is the normal state of multi-modal clinical data: the same reality is recorded, in different formats, by different specialties, against different reference points — and each version is, in its own context, "correct."

In Post 1, I argued that a research-ready cohort needs semantic consistency, temporal completeness, and computability. In Post 2, I described a multi-agent architecture (MAS) where a Data Governance Agent sits at the base, cleaning and standardizing data before any reasoning happens. Today I want to go deeper into that base layer — because this is where the real bottleneck lives, and where the data-level influence of our field will ultimately be won or lost.

Why Multi-modal Governance Is the Hard Part

Collecting data is easy. Governing it is not. Three structural problems make multi-modal colorectal cancer (CRC) data genuinely difficult:

1. Heterogeneity

The same patient generates tabular lab values, DICOM imaging, PDF pathology reports, operative videos, and free-text notes — four incompatible formats carrying three incompatible semantics.

A CEA value is a number. A circumferential resection margin is a measurement with a threshold. An MRI "suggests" venous invasion; pathology "confirms" it. Treating these as interchangeable rows in a spreadsheet is the original sin of clinical data science.

2. Conflict

When the same fact appears in three places and disagrees, what do you store?

Average them? Discard the outlier? Pick the newest? Each default decision silently corrupts downstream research. In our pre-standardization audit, source disagreement on key postoperative variables reached up to 30% — the exact gap between a real finding and noise.

3. Temporality

A tumor measured before neoadjuvant therapy is not the same tumor measured after it.

Every data point must be anchored to a patient timeline, or the cohort collapses into a cross-sectional snapshot that cannot support survival, recurrence, or treatment-response questions.

Four Engineering Primitives of Governance

We built the DACCA (DAtabase for Colorectal CAncer) cohort — now 6,000+ cases with 30-year follow-up, spanning over 50 dimensions of clinical, imaging, and molecular variables — around four non-negotiable primitives. These are not abstract principles; they are encoded rules that run on every incoming record.

The discipline here matters more than the tooling. A source hierarchy is worthless if a clinician did not define it. That is why, in our group, the variable dictionary and the conflict-resolution rules are written and versioned by surgeons — not by the IT team that designs the schema.

DACCA in Practice: The Data Governance Agent

This is where the architecture from Post 2 becomes concrete. Our Data Governance Agent executes the four primitives automatically:

• It reads a radiology report, extracts the TNM stage, and writes it to the structured DACCA field.

• It cross-checks that extraction against the pathology report.

• When they disagree, it does not silently pick one. It raises a "requires human adjudication" flag and routes the case to a clinician — exactly the conflict-resolution path a human would take, but systematic and auditable.

We have published the underlying philosophy of this disease-specific data engineering as a representative paper in the DACCA series:

Wang XD, et al. West China 30-year CRC data engineering: disease-specific digitalization and value-based care. Chinese Journal of Bases and Clinics in General Surgery. 2025.

That paper documents how a three-decade CRC cohort was transformed from a passive record repository into a governed, computable research platform — the same transformation I have argued for across this series. The DACCA program has now produced 20+ publications under the unified subtitle "real-world data from DACCA," establishing the dataset itself as a research asset rather than a storage artifact.

The Translator Between Data and Intelligence

Here is the thread that ties this post back to the first two.

A MAS is only as trustworthy as the layer beneath it. The Clinical Decision Agent and Follow-up Agent from Post 2 reason over data that the Governance Agent has already aligned, conflict-resolved, and made traceable. Remove the governance layer, and the intelligent agents are reasoning over three conflicting versions of the same tumor — confidently, and wrongly.

This is why I keep returning to one claim: the governance layer is the translator between raw clinical reality and artificial intelligence. It is unglamorous. It will never win an image-generation contest. But it is the prerequisite for everything else.

And the payoff is measurable. Once the data is governed, it becomes a substrate for genuine prediction. In our group's work on time-to-event machine learning for CRC survival:

Yang X, Qiu H, Wang XD, et al. Predicting Colorectal Cancer Survival Using Time-to-Event Machine Learning: A Retrospective Study. J Med Internet Res. 2023;25:e44417. DOI: 10.2196/44417

That model was only possible because the training cohort was built on governed, temporally complete, conflict-resolved data — not on the raw, heterogeneous records as they exist at the point of care. Good AI is downstream of good governance.

Why This Is the Data-Level Battleground

I want to be explicit about the strategic stakes, because this is where I believe our community should concentrate its effort.

There is no shortage of colorectal cancer AI demos. There is a severe shortage of governed multi-modal cohorts that those demos could honestly train on. The bottleneck is not algorithmic. It is the governance layer — the unglamorous work of defining source hierarchies, encoding uncertainty, and enforcing traceability across imaging, pathology, and text.

For those of us building medical data element engineering programs — turning clinical data into assets that are standardized, compliant, and reusable — this layer is the product. A cohort that cannot be governed cannot be an asset. Governance is the moment raw data becomes research capital.

What's Next

In the next post, I will move from the governance layer to what it unlocks: data-driven clinical research on the DACCA cohort. What specific, non-obvious clinical questions can 6,000+ governed cases with 30-year follow-up actually answer — and how do we avoid the trap of "big data, small insight"?

Selected Readings

If you would like to go deeper into the work behind this post, these publications are a good starting point:

1. Wang X, Qiu H, Wang XD, et al. Predicting Colorectal Cancer Survival Using Time-to-Event Machine Learning: A Retrospective Study. J Med Internet Res. 2023;25:e44417. DOI: 10.2196/44417 — an example of what governed, temporally complete CRC data makes possible.

1. Wang XD, et al. 华西30年CRC数据工程：专病数字化与价值医疗 [West China 30-Year CRC Data Engineering: Disease-Specific Digitalization and Value-Based Care]. Chinese Journal of Bases and Clinics in General Surgery. 2025 — the DACCA data-engineering philosophy, representative of a 20+ paper series.

Join the Conversation

I'd like to hear how others are tackling this:

• What source hierarchy does your group use when imaging and pathology disagree on staging?

• Do you encode uncertainty as a first-class variable, or treat "unknown" as a simple null?

• For those publishing clinical AI: how much of your pipeline is data governance versus model development?

Comment below or reach out directly. The governance layer is where colorectal cancer data science will either mature or stall — and we should be comparing notes openly.

The views expressed in this post are the author's own and do not represent the position of any affiliated institution.