Planning, defining success, and working with secondary data felt like separate problems at first. But they turned out to be the groundwork for two things I have to do throughout the project: assess it while it's running, and evaluate it once testing is done.
Planning against a moving target
Data imputation for clinical fields isn't a settled area. New methods and evaluation approaches keep appearing. That makes committing to a fixed plan feel slightly uncomfortable.
The Phelps, Fisher and Ellis (2007) planning chapter helped here. Experimentation is iterative, not linear. So milestone dates carry a buffer, not a promise. That's also the reasoning behind running the project in one-week Scrumban sprints (Ladas, 2009). Each sprint has a Review step, which weighs progress against my success criteria. Then an Adapt step decides whether to keep, adjust, or drop a method before the next sprint starts.
I track all of this in ClickUp: a Gantt chart for the overall milestones, a Scrumban board for the weekly sprint cycle, and custom fields as a running decision log — what was tried, what the result was, and whether it was kept or dropped. This is what ongoing assessment actually looks like. Not a single check at the end. Deviations get flagged as they happen.
Defining success when the field itself hasn't settled
If the plan has to stay flexible, the definition of "done" needs to be just as deliberate. Otherwise there's no way to tell whether an adjustment is progress or just drift.
Wingate's (2014) distinction between verification and validation gave me that anchor. Success means the imputation is built right — passing the RMSE, distribution, and significance thresholds. It also has to be the right thing — meeting the F1/AUC floor that shows it actually improves readmission prediction. Conflating the two would let me claim success on a technicality. A statistically plausible imputed value that does nothing useful downstream isn't a win.
That's why evaluation happens in two layers, not one: a distribution/significance test, and a classification-performance check. Both are judged together, never either alone.
Secondary data as the anchor
None of this planning happens in a vacuum. It's all built on a dataset I didn't collect myself. The Diabetes 130-US Hospitals dataset (Strack et al., 2014) is secondary data — 101,766 patient admissions across 130 US hospitals, gathered for hospital administration purposes, not for my research question.
That gap matters. Fields like medical specialty and payer code sit in a genuine judgement zone — roughly half missing, with no bright-line rule for impute versus exclude. Their missingness pattern was shaped by whatever administrative process generated it, not by anything I control or fully understand. Using secondary data responsibly means logging the reasoning behind each field's decision individually, rather than applying one blanket rule across the dataset.
Takeaway
Planning in a field that's still moving isn't a weakness to plan around. It's why assessment has to be continuous, not a single end-of-project check. And it's why evaluation has to test both that the work was built right, and that it's the right thing.
References
Ladas, C. (2009) 'Scrumban: Essays on Kanban Systems for Lean Software Development'. Available at: https://www.semanticscholar.org/paper/Scrumban%3A-Essays-on-Kanban-Systems-for-Lean-Ladas/09b6383d58f385c25580b095c32ca5246b1d1a84 (Accessed: 28 September 2026).
Phelps, R., Fisher, K. and Ellis, A.H. (2007) Organizing and Managing Your Research: A Practical Guide for Postgraduates. London, United Kingdom: SAGE Publications, Limited. Available at: http://ebookcentral.proquest.com/lib/sunderland/detail.action?docID=354865 (Accessed: 17 September 2026).
Strack, B. et al. (2014) 'Impact of HbA1c Measurement on Hospital Readmission Rates: Analysis of 70,000 Clinical Database Patient Records', BioMed Research International, 2014, p. 781670. Available at: https://doi.org/10.1155/2014/781670.
Wingate, L.M. (2014) Project Management for Research and Development: Guiding Innovation for Positive R&D Outcomes. 1st edn. Boca Raton: Auerbach Publications (Best Practices and Advances in Program Management Series). Available at: https://doi.org/10.1201/b17241.
Top comments (0)