In the Pathology Nucleus oncology database, which is built from The Cancer Genome Atlas (TCGA), the specific surgical procedure was recovered from the original pathology reports for 1,930 of 9,202 patient cases (21%). For 1,318 of these cases, the public structured data recorded no surgery at all. For the remaining 612, it recorded only the non-specific entry "Surgery, NOS". In the colon, prostate and kidney cancer cohorts, surgery was almost entirely absent from the public treatment records, although every case in those cohorts was a resection specimen.
Why apparently missing fields matter
Public datasets such as TCGA are widely used for model development because they are large and freely available. Their structured fields are, however, often incomplete, inconsistently coded or poorly linked to the slides they describe. A field recorded as blank is usually treated as missing, and the case is then either excluded or analysed without the variable. Where the information exists in the source report, both choices discard usable data. Exclusion can also bias a cohort towards the cases that happened to be coded well, which in TCGA differ systematically by cancer type.
Completeness and correctness are separate problems
Recovering a value makes a dataset more complete. It does not, by itself, make the dataset more accurate, because an extracted value can be wrong. Automated extraction from free-text reports works well for fields that are stated explicitly and consistently, such as the name of a surgical procedure. It is less reliable for fields that require interpretation, such as tumour size when several measurements are reported, margin status, or a diagnosis revised in an addendum.
For this reason, automatically extracted values in the Pathology Nucleus database are treated as provisional until checked. Values for fields that are diagnostically material or ambiguous are verified by a consultant pathologist, and the source of each value is recorded, so that a user can see whether it came from the public structured data, from automated extraction or from pathologist review.
Implications for teams using public data
A field recorded as missing should be checked against the source reports before a case is excluded or a variable abandoned, particularly where the missing values are concentrated in particular cancer types. Recovered values should then be verified to a standard that matches their intended use, with consultant review reserved for the fields where an extraction error would change a diagnosis, a stage or a cohort assignment.