Kaplan-Meier Curves

The survival curve you see everywhere -- and how to actually read it

The Setup

A paper reports "5-year primary patency was 72% for bypass vs 54% for endovascular repair." Below the text is a curve that slopes downward. Most readers glance at it, note which line is higher, and move on.

That is not reading a Kaplan-Meier curve. That is looking at a picture.

Kaplan-Meier curves are the most common way to present time-to-event data in surgical research. If you cannot read one critically, you are outsourcing your judgment to whoever drew it.

Anatomy of a KM Curve

Every Kaplan-Meier curve encodes the same information. Here is how to decode it.

X-axis Time since the index event (surgery, diagnosis, enrollment)

Y-axis Proportion still event-free (starts at 1.0 or 100%)

Steps down Each drop = one or more events occurred at that time point

Tick marks Censored observations -- patients lost to follow-up or study ended

Number at risk The table below the curve showing how many patients remain at each time point

Why It Steps Instead of Slopes

The curve is flat between events because the estimated survival probability does not change until someone actually has an event. Each step down recalculates the conditional probability of surviving past that time point, given that you made it that far.

The number-at-risk table is not decorative. If the curve shows 80% survival at 5 years but only 3 patients are still being followed, that estimate is nearly meaningless.

Reading KM Curves Like a Surgeon

Five things to check every time you see a Kaplan-Meier curve.

1. Check the Y-axis. Does it start at 100% or is it truncated? A Y-axis starting at 50% exaggerates differences between curves.
2. Check the number at risk. If fewer than 10-15 patients remain at a time point, the tail of the curve is unreliable. The confidence interval is wide even if the line looks steady.
3. Look at the separation. Do the curves diverge early, late, or never? Early separation suggests an immediate treatment effect. Late separation may reflect survivorship bias or delayed benefit.
4. Check for crossing. If curves cross, a single hazard ratio is misleading because the relative risk is not constant over time. This violates the proportional hazards assumption.
5. Look at the censoring pattern. Heavy censoring (many tick marks) means many patients were lost or the study ended before their outcome was observed. If censoring is not random, the curve is biased.

The Tail Problem

Journal reviewers now routinely require number-at-risk tables. If you are reviewing a manuscript that lacks one, that is a red flag. The tail of the curve is where authors can hide unreliable estimates.

Common Misinterpretations

Things people get wrong about Kaplan-Meier curves -- even at conferences.

"The 5-year survival is 72%, so 72% of patients are alive at 5 years."

Not necessarily. The KM estimate is based on conditional probabilities at each event time. If many patients were censored before 5 years, the estimate reflects those who were still being followed, not the entire cohort. Heavy censoring inflates the apparent survival.

"The curves separate, so one treatment is better."

Visual separation is not statistical significance. You need the confidence intervals and the log-rank test. Two curves can look separated by chance, especially with small sample sizes.

"Median survival was 3 years."

Only meaningful if the curve crosses 50%. If fewer than half the patients had the event, the median survival is undefined. Authors sometimes extrapolate it anyway, which is wrong.

"We can compare the curves at specific time points."

Cherry-picking time points is misleading. Comparing survival at 1 year vs 3 years vs 5 years can tell different stories depending on which you choose. The log-rank test evaluates the entire curve, not snapshots.

The Kaplan-Meier method makes one critical assumption: censoring is non-informative. If sicker patients drop out (informative censoring), the curve overestimates survival.

Exercise: Interpreting Kaplan-Meier Curves

Read each scenario and choose the best interpretation.

Question 1 of 8

Lesson Complete!

0

Total Points Earned

Exercise (0/8 correct) +0 pts

Perfect Score Bonus +50 pts

Lesson Completed +100 pts

✓

Kaplan-Meier Curves

Module 4 - Lesson 1 complete

Key Takeaways

X-axis: Time. Y-axis: Proportion event-free. Steps = events. Ticks = censored.
Number at risk: Always check it. Thin tails mean unreliable estimates.
Censoring matters: If censoring is informative, the entire curve is biased.
Visual separation is not significance: You need a statistical test.
Median survival: Only defined if the curve crosses 50%.
Y-axis manipulation: Truncated axes exaggerate differences.

Next up: Censoring -- the invisible force that shapes every survival curve and can silently destroy your analysis.