BMI and ICU Recovery: A Biostatistics Study

We pulled patients matching our criteria and ended up with 31. That's what MIMIC-III gave us after exclusions. But we didn't run a power analysis first to determine the minimum sample size needed to detect a clinically meaningful effect. A post-hoc power calculation showed that detecting a moderate correlation (r = 0.3) at alpha = 0.05 with 80% power requires roughly 85 patients. We had 31. The study was underpowered from the start, and we should have either broadened our inclusion criteria, included additional surgical categories, or acknowledged the power limitation in our study design rather than discovering it in our discussion section.

MIMIC-III records height and weight at ICU admission. We computed BMI from that single data point. But BMI at the moment of ICU entry may not reflect the patient's chronic weight status. Post-surgical fluid retention, edema, and perioperative weight changes all affect the number. A patient categorized as "obese" at ICU admission might have been "overweight" the week before surgery. We should have discussed this measurement validity issue more prominently. In retrospective database studies, you use what's available, but acknowledging the gap between what you measure and what you actually want to measure is important.

BMI doesn't exist in isolation. Age, comorbidities, surgical complexity, and pre-operative health status all affect ICU length of stay. Our analysis tested the bivariate relationship between BMI and LOS without adjusting for confounders. A multivariable regression or stratified analysis would have been more informative. We flagged this as a limitation, but the better approach would have been to include at least age and Charlson comorbidity index as covariates from the start. The simple bivariate design was appropriate for a semester project, but it limits what the null result can actually tell us.

Because null results are honest results. The temptation in a portfolio is to only show projects where everything worked and the numbers were impressive. But in clinical research and medical device validation, the ability to report what the data actually shows, even when it contradicts your hypothesis, is more important than flashy results. I included this project because it demonstrates that I can design a study, execute the analysis, and report findings without inflating them. If I'm going to work in a regulated industry where data integrity matters, this is the kind of work that proves I take it seriously.

Clinical outcome analysis is a core competency in medical device post-market surveillance and clinical studies. Understanding how to query clinical databases, apply appropriate statistical tests, handle missing data, and interpret results is directly applicable to clinical investigations under 21 CFR 812, post-market clinical follow-up studies, and real-world evidence submissions. The specific question (BMI vs. ICU LOS) is less important than the methodology: retrospective cohort design, non-parametric testing for non-normal distributions, honest reporting of limitations, and cross-validation of results between Python and R.

First, I'd expand the cohort to at least 100 patients by broadening surgical categories or using MIMIC-IV (larger dataset, more recent). Second, I'd include confounders: a multivariable logistic regression with age, Charlson comorbidity index, and surgical complexity as covariates. Third, I'd switch from WHO BMI categories to continuous BMI with restricted cubic splines to capture non-linear relationships. The current study answers a narrow question with limited data. A larger study could answer the more useful clinical question: after adjusting for known risk factors, does BMI independently predict prolonged ICU stay?