How upstream gaps in documentation, translation, and follow-up cascaded into limited options—and why I now see healthcare as an information system.
My mother was diagnosed with cancer in 2019. The diagnosis did not feel like a single moment—it felt like the final collapse of a long chain of unclear documentation, ambiguous follow-ups, and silent errors that accumulated over years.
For a long time, I blamed individual decisions. Only later did I realize it was, in many ways, a system problem.
Before coming to the United States, my mother visited a small community hospital in China. The record they produced was sparse—more of a brief note than a structured history. There were no clearly defined fields, no summary of risk, no longitudinal context. The doctor handed her a slip of paper and told her to “follow up if symptoms continue.”
Like many patients in low-resource communities, my mother assumed that “no call” meant “no problem.” The hospital never reached out. No one explained the significance of the abnormal findings. No flags were raised; no tracking system followed her case.
Years later, when she was finally seen at UCSF, physicians reconstructed her history and pieced together evidence that should have triggered closer follow-up much earlier. The signal was there—but it had not been represented in a way that the system could act on.
When she eventually received a formal diagnosis, her tumor was around four centimeters. It had not yet spread. I asked the local doctor whether surgery was still possible.
He said no. He said the window had passed. He said a second opinion would not change anything, because “the result will be the same everywhere.”
I believed him. We did not seek another opinion at that time. I still think about that moment.
Only later—after she transferred to UCSF—did we learn how long her symptoms had been present, and how vague documentation and follow-up had been. The lack of structure in her early medical records made the progression almost invisible from a systems point of view.
Our first appointment at UCSF was unlike anything I had seen before. A nurse spent more than two hours reconstructing my mother’s entire medical history—symptoms, tests, surgeries, medications, and small details from over a decade.
It felt like watching a broken record slowly being rebuilt: scattered fragments turning into a coherent timeline. In that reconstruction, we finally saw what had gone wrong:
· Early symptoms were documented vaguely or not at all.
· Follow-up was recommended but not clearly explained.
· No outreach was made when she did not return.
· Translation gaps led to partial or inaccurate communication.
· Key clinical details were missing or compressed in prior notes.
The UCSF team did not just treat her—they repaired the information structure that previous encounters had failed to create.
Throughout her treatment, I attended nearly every appointment as her interpreter. Even when official interpreters were present, I found myself correcting them often:
· Symptoms were simplified or mistranslated.
· Severity was understated.
· Pain descriptors were flattened into generic English terms.
· Cultural nuances were quietly removed.
Sometimes the interpreter was impatient. Sometimes my mother felt the impatience and stopped asking questions. These moments were more than uncomfortable interactions—they were points where data was quietly distorted.
In the EHR, clinicians documented what they believed they heard, filtered through the interpreter’s choices. The final note might not match what my mother actually meant. This is what I now think of as semantic drift: a shift in meaning as language crosses linguistic and cultural boundaries.
Semantic drift becomes documentation variability. Documentation variability becomes data noise. Data noise becomes decision risk.
In applied mathematics and numerical analysis, we talk about error propagation: how small deviations at one step accumulate and magnify downstream. In healthcare, I watched error propagation unfold as lived reality.
A vague note becomes an ambiguous diagnosis. An ambiguous diagnosis leads to a delayed referral. A delayed referral narrows the treatment window. A narrowed window limits the available options.
Data is not just an abstract input to a model. It is a sequence of decisions and non-decisions that shape real lives.
When I later studied data science and modeling, I started to see my mother’s case through a structural lens:
· Upstream representation determines downstream possibility.
· No model can recover information that was never documented.
· No algorithm can fix gaps created by translation drift and inconsistent notation.
· Multilingual patients often appear in the data with lower-quality, noisier records.
My mother’s story is not just personal; it is a case study in information loss. It is the reason I am interested in documentation variability, multilingual data inequity, and the way clinical information flows through systems with different levels of structure and support.
This experience is why I do not see health data as neutral. I see it as the result of:
· Who has time to document carefully.
· Who has access to interpreters—and how good those interpreters are.
· Which hospitals have follow-up systems.
· Which patients feel empowered to ask questions.
· How systems respond when someone does not return.
For me, health data science is not only about building better models. It is about asking:
· How is information created in the first place?
· Where does it drift, compress, or disappear?
· How do multilingual and low-resource patients get represented in data?
· What kinds of structure would have changed my mother’s trajectory?
This essay is not simply a story—it is the starting point of my research agenda. It is why I want to work on clinical documentation, multilingual EHR, and information flow mapping: to reduce information loss and support more equitable care.