Mortality

I computed the Kaplan-Meier estimates of survival probabilities as a function of time (in hours, where 0 is ‘at the scene’, and then we have 24, 48, and 720, i.e., 30 days) by category (Purple, Red, Amber, Yellow, Green, Not Categorised) and year (2016, 2017). The aim was to assess differences in survival within and across years.

The following plots describe the probability of survival by category in the month of January 2016 and 2017.

Within both years, survival for purple-coded patients appears markedly lower with respect to all other causes. Differences of practical significance between years seem to exist only for the purple-coded patients.

The following plots show estimated survivals by year within each category (the y-axis is truncated for all but the purple category for better readability).

There seems to be a considerable (~20%) increase in survival for purple-coded patients from 2016 to 2017 which is constant over time from time 0 (when the ambulance arrives at the scene). I believe this wide a gap may be imputable to the modifications of the allocation system. Possibly, more people which is less at risk of immediate death is being coded as purple ‘just in case’, which may in turn cause an artificial inflation of the survival probability: this needs to be tackled by examining the accuracy of allocation, i.e., how many individuals were correctly coded as purple or not in front of an immediate life threatening condition or absence thereof.

Here are the tables that Tony produced to investigate this.

2016 Confirmed purple Not purple
Purple 143 218
Other 8533 14690
2017 Confirmed purple Not purple
Purple 285 392
Other 9617 15122

In order to fill these tables, we worked under the assumption that those dying at the scene and those admitted immediately upon arrival at the ospital should be treated as confirmed purple, whereas those not conveyed or discharged upon arrival at the hospital should be considered non purple. If this sounds reasonable, then we observe a slight (but seemingly statistically significant, Tony to check on the confidence interval calculation) decrease in overall accuracy from 2016 to 2017, as defined by the correct number of classifications over the total: \(\frac{143+14690}{143+218+14690+8533} \approx 63\%\) in 2016 against \(\frac{285+15122}{285+392+9617+15122} \approx 61\%\) in 2017. We are looking at data for one month per year only, so we should be super careful about drawing conclusions from here, but if we also look at the rate of false positives (classified as purple, when they indeed were not), we go from \(\frac{218}{143+218} \approx 60\%\) in 2016 to \(\frac{392}{285/392} \approx 58\%\) in 2017, a difference of small magnitude, which could suggest that not so many individuals were conservatively allocate to purple (as I tentatively suggested looking at the survival curves above).

Overall, it looks like only Purple and Amber categories benefitted of the change of allocation criteria (their survival curves are higher for 2017 than for 2016); for the remaining codes, either no sensible difference or a very moderate decrease in survival is observed. Further investigation of the category-specific survivals would require an expertise such as David’s, and may provide additional insight into what happened.

Analyses re-run by condition codes

Survival analysis by condition.

Survival by condition within January 2016 and 2017 has pretty much the same structure. Cardiac arrest (accounting for most of the deaths under Purple code), presents a markedly lower survival with respect to all other conditions, which is not surprising. The only evident differences between the two years seem to be pertaining cardiac arrests (better survival in 2017) and breathing difficulties (worse survival in 2017). We explore this by constrasting the survival by year, for each condition. Note that all conditions but Cardiac Arrest have had the y-axis rescaled for better readability, so do not attempt graphical comparison across plots.

As suspected following inspection of the yearly plots, cardiac arrest presents an overall better survival in 2017 than it did in 2016 (\(\approx 10\%\) difference throughout). Breathing difficulty, on the other hand, seems to have worsened, with a 3% to 6% decrease in survival: it also seems that the gap widens as time passes. No remarkable differences seem to be present for Stroke and Falls (max difference in estimated survival probabilities \(<1\%\) at any point).