Pancreatic cancer phenotypes

The question

Pancreatic cancer has one of the worst survival rates of any cancer, largely because it is usually found late. If we understood its full set of signs and features (its phenotype) a little better, we might spot the people who need screening earlier.

Each source sees a different slice of a disease. Clinicians write down what matters clinically. Patients describe what they feel. Published research records what has already been established. My dissertation asked what each of these three sources contributes when you mine them the same way, and where they disagree.

What I found

• Clinical notes surfaced by far the most pancreatic cancer phenotypes: 595, against 136 from the literature and 21 from social media.
• Each source leaned a different way. Social media phenotypes clustered around neoplasms, the literature around neoplasms and the digestive system, and clinical notes most heavily in the genitourinary category.
• Patients and clinicians describe pain differently. Pain was more strongly associated with disease in patient language (average score 0.108) than in clinical notes (0.044). That fits other research showing that “discomfort” often goes unrecorded.
• Social media raised questions worth asking. Behavioural phenotypes such as aggression appeared only in patient language. Is the psychological side of a hard diagnosis being under-recorded?

How it worked

1. Collect. De-identified hospital discharge summaries from MIMIC-IV (331,794 admissions), social media data supplied by the charity White Swan, and phenotypes mined from biomedical literature.
2. Annotate. Text was tagged against the Human Phenotype Ontology and Disease Ontology using the Komenti text-mining framework.
3. Generalise. Specific terms were rolled up to their parent concepts, so “abdominal pain” also counts as “pain”.
4. Score. Disease–phenotype pairs were scored by how often they co-occur (normalised pointwise mutual information).
5. Test. 2,000 Monte Carlo simulations per association produced p-values, then q-values corrected for multiple testing at 0.0005.
6. Compare. Phenotypes were grouped into 20 high-level categories and normalised, so sources of very different sizes could be compared fairly.

Limitations

These matter as much as the findings:

• The clinical data is from intensive and emergency care at one US hospital, so it over-represents very unwell patients.
• The three sources were processed with different upstream methods. I aligned the scoring, but some bias is likely to remain.
• Short posts behave differently from long documents. Tweets leave less room for loose associations than a 4,000-word paper.
• Clinician review was sobering. A clinician rated the phenotypes unique to social media at an average of 1.15 out of 4, between “unlikely” and “feasible”. These are hypotheses to investigate, not discoveries.

No restricted source data is published here. MIMIC-IV requires credentialed access, and the social media data was provided for research use.