Examining the fire risk in London dwellings using the London Fire Brigade Incident database

Analysis of the Fire Brigade's database of fires in London between 2009 and 2020 provided insight into the level of fire safety in the city and how it varies across different types of dwellings and different levels of protection. Regarding the number of fires, fatalities, and injuries, fire safety in London has significantly improved on average over these years. However, average trends cannot analyze catastrophic fires with multiple fatalities, like at Grenfell Tower in 2017, as these events are too rare to form a suitable sample size. Dwelling fires are the most lethal in London: despite accounting for only 28% of fires, they lead to 87% of fatalities and 83% of injuries. The odds of a dwelling fire becoming fatal in London fell from 1 in 174 in 2009 to 1 in 208 in 2019, a decrease of 16%. The total number of fires has decreased over this period, and the number of fires where an alarm was raised has increased, suggesting that the prevention and detection layers of fire safety have improved, while our analysis suggests that the level of protection from the compartmentation and evacuation layers has remained constant over time. An analysis of the different layers of fire protection suggests that compartmentation was the most impactful layer, with a failure in compartmentation increasing the odds of a fire being fatal by 1.5 to 5 times. Overall, this analysis shows that the fire hazard to Londoners in general is low and the lowest since 2009; however, there is still a threat that should not be understated.

i.e., occupied by households, excluding hotels, hostels and residential institutions," by the London Fire Brigade's recording system manual. 2e impact of different layers of fire protection was also investigated for fires in blocks of flats.
Between 2009 and 2020, the London Fire Brigade (LFB)-the busiest fire and rescue service in the UK and one of the largest in the world-responded to a total of 262 372 fires that resulted in at least 14 176 injuries and 647 fatalities.In 2020 alone, the LFB responded to 17 409 fires that resulted in at least 832 injuries and 32 fatalities.
Data from fire incidents is recorded in the LFB incident database as Fire brigades across the UK are required to collect data on fire incidents through the national Incident Recording System (IRS). 2 However, the LFB records additional information on top of this system and leads the way for fire and rescue services in terms of data collection and analysis.This information represents a rich source of data for understanding the level of fire risk faced by London citizens.The data and their analysis can help engineers, designers, the fire services, authorities, and policymakers to make informed decisions on how best to reduce the fire risk.

| Risk and layers of protection
The common definition of fire risk 3 (R) combines the probability that fire could occur (likelihood, L) with the negative consequences of fire (severity, S).This is expressed by Equation (1).Reducing fire risk therefore involves reducing the number of fires that occur each year, and reducing the harm caused by each fire.
The two elements of risk are part of how safety is implemented in a building via the layers of fire protection.These are, in their simplest form and in approximate chronological order, the following: prevention, detection, evacuation, compartmentation, suppression, and structural resistance.The layers aim to reduce the fire risk of a building, with five of the layers aiming to reduce fire severity, while prevention reduces fire likelihood.All layers of protection interact with one another and, in general, if one layer is breached another layer can compensate and mitigate the severity.There are also superlayers of protection, such as fire brigades and tools to ensure the presence of the layers of protection, such as building regulations and management procedures that act on multiple layers at once.The Swiss Cheese model of risk, first proposed in the context of more general risk analysis, 4 is applied in Figure 1 to the layers of fire protection in a building (in which both fire brigades and building regulations are referred to as superlayers in their capacity to act upon multiple layers).
The six main layers of fire protection are summarized below and are described in more detail elsewhere. 5Prevention works by understanding the causes of fire, increasing awareness of fire safety, ensuring electrical safety, and keeping ignition sources away from flammable items.
• Detection works by sensing the fire early through its smoke, light, or heat, and sounding the alarm or notifying occupants early enough that they have sufficient time to act while the fire is small.
The Swiss Cheese model of risk is applied here to the different layers of fire protection in a building.A hole in one layer of protection can be compensated by one of the other layers to mitigate the risk.Only when multiple layers are breached (the holes lining up) will a fire become severe.Superlayers, like fire brigades and building regulations, act on multiple layers at once.
• Evacuation works by providing measures to ensure the safe egress of the occupants to a place of safety, avoiding exposure to fire and smoke.It is of critical importance to life safety and requires clear routes of escape.
• Compartmentation works by restricting fire spread through the walls and ceilings of the building, creating "fire compartments" of limited size, and preventing the fire from becoming uncontrollable.
• Suppression works by applying active measures to extinguish or control the size of a fire.
• Structural resistance works by ensuring the building's structure can continue to withstand the building's mechanical load during a fire.
Note that these six layers of fire protection are only one way of conceptualizing the process of fire safety design and management.
For instance, Layers of Protection Analysis (LOPA) in industrial fire protection engineering uses a completely different approach to selecting layers of protection. 6

| Review of fire statistics
Research into fire incident statistics is vital in informing how to reduce fire risk, be that noting inadequate minimum requirements in building code, as found in one case study review 7 ; or the effectiveness of public health interventions to reduce the risk of fire injury and fatality for vulnerable members of society suggested by another study. 8so noted in previous studies were possible issues with these types of analyses, in that recorded fire service incident data can be innately skewed towards larger fire, as smaller fires not attended by the fire service are not included. 9Additionally, when considering how fire risk varies with dwelling type-such as low-, mid-, and high-rise flats compared with singly occupancy houses-one study 10 showed that despite residents in multistorey flats in England initially appearing no more likely to die or be injured than in any other dwelling; after normalizing the data by the estimated population living in each dwelling type, those residents living in flats were exposed to a greater likelihood of their building experiencing a fire and were more than twice as likely to die in such an event.
Furthermore, a number of other studies showed that fire risk in different dwellings were also affected by differing human behaviour.
For instance, high-rise dwellings were found to result in more complex behaviours than in single occupancy dwellings, for instance, people initially alerted by detection systems but not in direct contact with fire were found to react with delayed response and ambiguity. 11On the contrary, another study into human behaviour 12 found that in single occupancy housing, people tend to investigate the source of smoke or fire alarm in an attempt to put out the fire themselves rather than evacuate the dwelling.This demonstrated the significance of considering which factors were being accounted for in the analyses and the potential need to consider the impact of additional factors outside the statistics on fire risk.Another important consideration, when reflecting upon and comparing with previous research, is the environment in which the data were collected, and how different regulations and fire service practices may result in differing feature definitions resulting in comparisons becoming less straightforward. 13rther explored were the limitations of the previous research carried out, for instance, a number of studies 9,[14][15][16] used older data, be that due to the date the papers were published or in some cases, the accessibility of up-to-date data.In both cases, the trends and conclusions reached in said papers would need to be compared with those found using recent data in order to provide a more accurate picture on how relevant they remain.Moreover, limitations in the data regarding the extent or detail of the data acquired were also noted as limiting the analysis in a number of the evaluated studies, 14,17,18 this ranged from issues with lower quality data, small datasets to certain aspects having not been well-characterized.Consequently, this research is novel and significant in its use of the up-to-date and extensive LFB fire incident database to assess the level of the fire risk in London between 2009 and 2020.The results of which will allow for more informed decisions regarding the fire safety of Londoners and the buildings they inhabit.

| Dataset
Throughout this report, imagining the data to be formatted in a table, we will refer to each row of the table as an incident and each column of the table as a feature.Incidents refer to false alarms, fires, or special services ("non-fire incidents which require the attendance of an appliance or officer" 2 ).Features are properties of each particular incident, such as the number of injuries that resulted from the incident or the year in which the incident occurred.This terminology was adapted from Google's machine learning guidance documents. 19The analysis was carried out in the software R, primarily using the tidyverse suite of packages. 20e raw dataset was a subset of the London Fire Brigade's com- Descriptions of the particular features used in the research presented here are given in the appendix.In general, the features measured the time of the incident, the type of property, the severity of the fire (injuries, fatalities, and size of fire spread), and the effectiveness of four of the different layers of fire protection (detection, evacuation, compartmentation, and suppression).It is worth noting that for the incidents in this dataset, in each case the prevention layer had failed (as a fire had broken out), the detection layer had, to some degree, succeeded (as the fire brigade had been called), and the structural resilience layer had succeeded (as no structural collapse took place).

| Analysis methods
The average change in a feature, such as the total number of fires each year, over the 12-year period was estimated by using the gradient of an ordinary least-squares (OLS) line of best fit to the data over the entire period.Over long periods of time, a linear fit is usually not appropriate for predicting changing trends in time, as there will be many individual events that could influence the context of the trend.However, as a first approximation of the average change in number of fires/injuries/fatalities, etc., over this period, we believe that a linear fit appropriately captures the relevant trends.To calculate the percentage change in a feature between 2 years, we evaluated the line of best fit at those 2 years and used the percentage difference between these two values.
When assessing the average area damaged per fire each year, we used the variable AreaDamagedByFlameSqM.This variable is a rough estimate of the horizontal area damaged by heat and flame in the fire, made by the fire officers who arrived on site.The variable is stored as categories such as 0, less than 5 m 2 , 6-10 m 2 , 11-20 m 2 , etc.To calculate the average area, we had to convert these categories to a numerical value.We did this by taking the midpoint of each category, with the exceptions of the categories 0, which remained as 0, and > 10 000 m 2 , which was stored as 10 000.We used the same method to calculate the average evacuation time, which was also stored in categories of 0 min, 0-5 min, 5-30 min, etc.
The effectiveness of different layers of fire protection was measured using the odds ratio (OR), which is a common approach to evaluating the impact of Boolean predictor variables on Boolean outcomes. 21This is defined using the following equation: Where: The odds ratio represents the decrease in the likelihood of a fire being fatal based on a layer being present.For instance, an odds ratio of 0.5 would mean that a fire is half as likely to be fatal given that the layer is present.This equation was used to evaluate the impact of the following variables (described in the Appendix A) on whether a fire was fatal or not (WasFireFatal): • whether the automatic alarm system is activated (AlarmRaised) • whether evacuation was delayed, given that evacuation took place (EvacNoDelay) • whether compartmentation was successful (CompartmentationSuccess) • whether suppression was successful (SuppressionSuccess) Using Equation ( 2) assumes that these variables are independent of each other.This can be considered a reasonable assumption, based on the following arguments.Firstly, although the alarm being raised will impact whether an evacuation takes place, it will not affect whether the evacuation was delayed (given that an evacuation is taking place).Secondly, CompartmentationSuccess and SuppressionSuccess could be related, in that a successful suppression system will increase the likelihood that the compartmentation stopped/checked spread; however, there were such a small number of incidents where suppression systems were present that the analysis was done excluding these incidents.The success of compartmentation would then reasonably be considered independent from whether the alarm was raised and whether the evacuation was successful.The Odds Ratio and the associated confidence intervals were calculated using the epitools package 22 in R.
For many features, there were a significant number of incidents where that feature was not recorded, or unknown.These missing variables were imputed using the Multiple Imputation of Chained Equations (MICE) methodology, 23 which is commonly used in medical statistics.The assumption of this process is that the missingness of the data can either be predicted by the features within the dataset (for instance, the time the incident took place), or is distributed completely randomly, that is, the missingness is not correlated with some further feature outside the dataset (for instance, the experience of the firefighter recording the data).This assumption may not be completely accurate, but it is the best approach available under the circumstances.
The reliability of the different features was also considered.It was relayed by the LFB that the fire protection layer features were less reliable than the features describing the severity of the fire incidents, after some contradictions were discovered (for instance compartmentation was sometimes labelled as being successful, even though the fire was labelled as spreading over more than one floor of the building).While the fire and rescue service incident recording system has a number of quality assurance measures in place 24 it is not possible to verify the accuracy of every incident.Similarly, some definitions (such as the point at which people can be considered to have safely evacuated for the purposes of measuring evacuation time or delay) may not be entirely consistent throughout the data.Therefore the conclusions presented in this study are considered indicative, and further studies should be done to assess the reliability of the data available.

| Frequency of fires
The London Fire Brigade is the busiest fire brigade in the country, as evidenced by the approximately 1.3 million incidents it recorded between 1 January 2009 and 31 December 2020.Of these incidents, 396 620 (31%) were "special services" (such as traffic collisions or flooding), 638 943 (49%) were false alarms (incidents "where the LFB attends a location believing there to be a fire incident, but on arrival discovers that no such incident exists, or existed." 2 ), and 262 372 (20%) were actual fires.These results come from the data available on the London Datastore website 25 ; however, the detailed dataset received for this research was slightly different.The dataset analyzed here included 901 686 incidents, 1478 (0.2%) of which were special services, 638 225 (71%) of which were false alarms, and 261 983 (29%) of which were fires.These are only 718 (0.11%) fewer false alarms and 389 (0.15%) fewer fires than the complete dataset, so this should not impact the findings of this research.These proportions have remained relatively steady year-on-year, as shown by the constant gap between the lines of best fit in Figure 2A.Calls to the fire brigade tend to peak at 7 pm, perhaps when most people are preparing dinner or spending time at home, and recede at 5 am, perhaps because most people are asleep at this time.These trends can be seen in Figure 2B.A recent report from the UK Home Office 26 on the economic and social costs of fires in England suggests that fires where the fire brigade attend (all the fires in this dataset) represent only a small percentage of the true number of household fires.However, fires where the fire brigade does not attend are generally minor fires that do not cause injury or significant damage.Therefore this should not impact the results presented here.
The number of fires and false alarms in London has been steadily decreasing over the last 12 years, with an average decrease of 975 fires per year, or a 39% decrease since 2009.This downward trend seems to be consistent with other fire brigades across England. 1 However, there was only a 28% decrease in the number of dwelling fires per year, meaning even more fires were prevented outside of residential properties, which was similar to the decrease in other types of buildings.The larger decrease in overall fires comes from a significant drop in outdoor fires.This is due to a change in April 2012 to how these fires were defined in the Incident Recording System, which also accounts for the sudden drop in fires between 2011 and 2012 (this sudden drop is also present in Figure 3A).In fact, when only considering the data from 2012 to 2020, outdoor fires are the only category of fire that increased in frequency, by roughly 47 fires per year (though the linear fit was not statistically significant in this case).Possible reasons for the overall decreasing trend in building fires are improvements to the prevention layer of fire protection, such as better awareness of fire safety, fewer ignition sources safer consumer products, improved regulations, and other advances in fire safety.
These trends remain positive when viewed in light of the city's changing size over this time.Over the 12-year period, there has been a steady increase in London's population, from 7.9 million in 2009 to 9.3 million in 2020-an average increase of 117 000 more Londoners every year. 27Similarly, the total number of dwellings in London has increased from approximately 3.3 million in 2009 to 3.7 million in 2020-an average increase of 33 000 new dwellings each year 28 (meaning, there was on average 1 dwelling built for every 3.5 new Londoners).This means that the decrease in the number of dwelling fires proportional to the number of people and dwellings was even higher, with an average decrease of 49% fewer fires per person per year (considering all types of fires) and 35% fewer dwelling fires per dwelling per year over this period (see Figure 3).This means that in 2009 the proportion of London dwellings that experienced a fire was 1 in 473, but by 2019, this had fallen to 1 in 680.Looking at different categories of dwellings in Figure 7, we can see that the number of injuries and fatalities is decreasing in a similar way to the total number of dwelling fires.The comparatively shallow decrease in the number of fire injuries and fatalities in 10+ story flats is mainly due to the Grenfell Tower fire.It seems that overall all types of properties have become safer over time.This does not mean that catastrophic fires are less likely in the future, but it seems that the general trend in London has been towards safer dwellings, either through engineering, education, effective firefighting, safer appliances, less smoking or more likely due to a combination of all of the above and all of the layers of fire protection.This is further supported by Figure 8, which shows the number of injuries and fatalities per dwelling fire per year.The number of injuries occurring on average in a dwelling fire decreased by 20% over this period, which supports the argument that not just the frequency but also the average severity of London dwelling fires is reducing.An indepth analysis of the reasons that fatalities per fire are not falling as quickly as injuries per fire should be the topic of future work.

| Injuries and fatalities
Figure 8B shows the large impact of The Grenfell Tower fire in the 2017 datapoint.More people died in Grenfell Tower than in any year in London due to fire.On average, and over the period 2009-2020, the number of dwelling fire fatalities per year has increased by 11% due to the impact of the Grenfell Tower tragedy.
Considering all the fires across this period (Figure 9), for all buildings below 10 storeys, the probability that one of these fires, once it had already started, would become injurious or fatal was similar across the dwelling types; at an approximately 1 in 9 chance that the fire would cause injury and a 1 in 172 chance that the fire would cause a fatality.However, fires in buildings of 10 storeys or more were less likely to cause an injury (a 1 in 11 chance) but more likely to cause a fatality (a 1 in 123 chance).It is unclear why fires in taller buildings would be less likely to cause injury while being more likely to be fatal.
Perhaps, there is a smaller proportion of serious fires in 10+ storey buildings, reducing the average number of fires that cause injuries, but the serious fires that do occur are more likely to lead to a fatality.It is also possible, given that these kinds of buildings are more often equipped with sprinklers, that there are fewer small fire incidents where the fire brigade attends, as the fire would be extinguished before needing to call the fire brigade, which would mean that fires in 10+ storey buildings in our dataset would be on average more severe.This trend could be something to investigate in further detail in future research.Overall, the odds of a dwelling fire becoming fatal fell from 1 in 174 in 2009 to 1 in 208 in 2020, a decrease of 16%.

| Catastrophic fires
The National Fire Protection Association (NFPA) keeps track of residential fires in USA including those where 5 or more people are killed, which they define as catastrophic fires, a useful definition that does not exist in UK fire statistics reporting.A recent report from the NFPA 29 showed that the number of catastrophic residential fires in the USA is decreasing slowly since 1988, with 190 such fires between 2009 and 2019.Figure 10 shows that there were also 3 catastrophic fires in London in the same period.This suggests that such large tragedies are thankfully rare occurrences (note that scaled by population, catastrophic fires are roughly 5-6 times more frequent in London than the US, though the population distributions are significantly different, and it is hard to draw consistent trends from such infrequent data).However, the reason for bringing up this definition of catastrophic fires is not to present a detailed comparison between London and US fire statistics (which is outside the scope of this study) but to  Neasden in 2011, 32 and the Grenfell Tower fire in 2017. 33There are some parallels that can be drawn between the fires.Both the Grenfell fire and the fire in Neasden originated from a faulty freezer, and both Lakanal house and Grenfell tower had flammable cladding that contributed to the spread of the fire.Twenty-four dwelling fires during this period caused 2 fatalities.These fires were in a variety of property types and were quite different in cause and size.However, only in 2 of these fires was the property's compartmentation successful, and none of them were in dwellings with active suppression systems.
There were no fatalities from any dwelling fire in the database where all the layers of protection were recorded as working as intended.

| Property damage
In addition to people's life and health, fire also damages the property.
Fire damage is related to property protection more than life safety, and is another negative consequence to consider in fire safety.The area damaged is defined within the incident database as the total floor or ceiling area (whichever was larger) that was damaged by heat.The data shows that once a fire occurs, the probability that it will damage more than the area of a single compartment (about 100 m 2 ) is 0.0035 or 1 in 284.The average area damaged in a dwelling fire during this period was about 4.4 m 2 .This number stayed roughly the same from 2009-2020, which suggests that the effectiveness of compartmentation has remained the same over the period.This also suggests that the continuous reduction in injuries and fatalities seen in Figure 8 is not due to further improvements in compartmentation, but due to improvements in other areas.Considering the layers of fire protection mentioned in the introduction, this suggests that improvements over this period have been to prevention, detection, and evacuation, as opposed to compartmentation or suppression (as improvements in these last two should have also reduced the average area damaged in fires over this period).

| Trends and impact of the layers of fire protection
Having discussed trends in dwelling fire frequency and severity over time, it would make sense to understand the impact of the different layers of fire protection on dwelling fire severity during this period (2009-2020).Information on the different layers in different property types is summarized in Table 1.Perhaps surprisingly, the number of fires where an alarm was successfully raised was less than 40% on average for all dwelling types.This is roughly 45% lower than the percentage of fires where an alarm was recorded as being present (between 56% and 64% of fires across the different dwelling types).
This suggests that an alarm is only successfully raised in a fire about half the time.However, this could be because these fires were detected in some other way and the fire brigade was notified before an alarm was raised.These data are also significantly different than the results of the English Housing Survey 2013-14, 34 which reported that approximately 89% of households had a working smoke alarm.The percentage of fires where compartmentation controlled spread was similar for blocks of flats (73-77%), regardless of height, while single occupancy houses had a lower frequency of successful compartmentation (47%).This is understandable, as compartmentation is not generally as important to life safety in a single occupancy house, where everyone is expected to evacuate as soon as a fire is detected, as opposed to blocks of flats, where the flat of fire origin is expected to evacuate, but occupants of other flats are expected to be able to remain in place if they wish (in the UK, this is known as a "stay put" evacuation strategy).The frequency of fires where compartmentation-controlled spread remained relatively constant from 2009 to 2020.
Active safety systems, including smoke control systems and automatic suppression systems, are not as common as other layers of safety in London.For context, UK fire safety guidance 35 until 2020 only specifically recommended residential sprinklers in blocks of flats over 30 m (roughly 10 storeys).In 2020, this was reduced to 11 m (roughly 4 storeys).Despite these recommendations, Table 1 shows that only 0.9% of dwelling fires were in buildings with active safety systems, and only 18% of these safety systems were automatic sprinkler systems.There were only 102 fires where the impact of the sprinkler system was recorded, and only in 53 of these did the system succeed in extinguishing or "containing or controlling" the fire.This means that of the small number of fires in buildings with sprinklers, only 52% of the fires were controlled.This is an unexpectedly low number that requires further investigation, although it is not entirely anomalous due to previous papers detailing the problems with sprinkler reliability, with one such study 36 finding the percentage of F I G U R E 1 1 Proportion of fires where at least one alarm was present (A, left) and was successfully raised (B, right).The line of best fit is for all types of dwellings combined (for clarity, given that the trend is similar for all dwelling types).
sprinkler systems that failed to carry out their function was relatively large.However, it could also be possible that these systems were so effective in controlling the fire in some cases that the fire brigade was never called in the first place.In this case, the low number of fires with automatic sprinkler systems in the incident database could be seen as a success, and the 52% figure could be underestimating their effectiveness.The English Housing Survey 2013-14 34 estimates that approximately 0.29% of homes in England were provided with sprinklers.In 2013 in London, 3 out of 6174 dwelling fires attended, or 0.05% were provided with sprinklers.This suggests that it is possible that sprinklers are helping to extinguish fires in incidents where the fire brigade did not attend.It is possible that this could be studied further in the future by combining incident data across England with more detailed survey data of the total number of sprinklered buildings; however, this is outside the scope of this study.Figure 12 also shows that the proportion of buildings 10 storeys or higher with sprinklers seems to have been growing with time (or at least the number of fires in buildings with sprinklers).It could be that in the future this safety layer will be more common in London.
It would also be of interest to estimate the relative impact of each of the different layers of protection in reducing fire hazards in dwelling fires.Equation ( 2) can be used to calculate the increase in the odds of a fire being fatal, given that a particular variable is true or not.In this case, we used this equation to calculate the impact of either compartmentation being breached, no alarms being raised, or evacuation being delayed (given that an evacuation has occurred).The assumptions behind this analysis are given in the methodology section.This analysis was restricted only to purpose-built blocks of flats (not individual dwelling houses), as these have similar design objectives and approaches to detection, evacuation, and compartmentation.
The impact of sprinklers was not included in this analysis, as fires involving sprinklers were too rare as to give any kind of significant result.Therefore the analysis considered only fires in purpose-built blocks of flats without sprinklers, and in the case of evacuation delay, only fires where evacuation took place.
Figure 13 shows the results of this analysis.The error bars represent the 95% confidence interval for the estimated increase in odds ratio for each variable.It seems compartmentation has the largest impact on the probability of a fire being fatal, increasing the likelihood of a fire becoming fatal by roughly 1.5 to 5 times.The failure of any alarms to raise news of the fire also increased the odds of a fire being fatal by roughly 1.5 to 3 times (though of course, for the fire to be in the database it has to have been detected eventually).Surprisingly, whether evacuation was delayed or not had almost no impact on the odds of a fire being fatal (given that an evacuation took place).This may be because, as mentioned above, blocks of flats in the UK are designed with the intent that only the flat of fire origin will need to evacuate during a fire.Confining the fire to the flat of fire origin is much more important to the success of this "stay put" evacuation strategy than ensuring that the evacuation is not delayed.In addition,  To this point, catastrophic fires with more than five fatalities are very rare in London (and more generally), with only three such fires out of 261 983 fires attended by the London Fire Brigade over the 12 years investigated.These three catastrophic fires had multiple breaches of the layers of protection.By contrast, there were no fatalities in any fire when all the layers of protection worked well.This suggests that the concept of safety via multiple layers is effective and that when they are breached the fire risk grows significantly.
The impact of dwelling property type was also considered and injuries and fatalities were found to have decreased across all property types.
However, the decrease has been slowest for buildings with 10+ storeys.
Once a fire has started, it is less likely to cause an injury in a 10+ storey building vs. other property types but more likely to cause a fatality.
While the number of fires, fatalities, and injuries has fallen, the average area damaged per fire has remained the same.This suggests that London has improved over this period much more in life safety than in property protection.It also suggests that the reduction in the lethality of the fires was likely not due to the layers of protection that contribute most to reduce the area of fire damage, that is, compartmentation and suppression.
Similar information to the London Fire Brigade data used in this report is also recorded by other fire brigades across the UK.It would therefore be interesting to extend the analysis to the whole of the country.This could also help to combat some of the small sample sizes of fires involving automatic suppression systems, for instance.It

APPENDIX A
The features received in the raw data from the London Fire Brigade, which were used in this research, are described in the table below.Whether there was any delay in evacuation (given that PersonsEvacuated was true), and if so why.14 different categories, e.g., "Evacuation, but no delay," "Building management poor," "Means of escape-not suitable," etc.

9.4
EvacuationTimeSpan Categorical (dependent on PersonsEvacuated) Estimated time taken for evacuation to be completed.Evacuation is defined in the documentation as "the direction of people from a dangerous place to somewhere safe."This includes people moving within the same building and does not necessarily mean the time for everyone to have left the building.9.5 plete database, received in tabular format (as a.csv file) from the London Fire Brigade directly.It contained details of 901 686 incidents between 1 January 2009 and 31 December 2020, including 1478 special services incidents (such as road traffic collisions or flooding) and 638 225 false alarms.Although the raw dataset contained 33 features, only features relating to the type of building, the damage caused by the fire (including deaths and injuries), and the different layers of fire protection were considered in this study.
Number of incidents where layer was present and fire was fatal • b = Number of incidents where layer was present and fire was not fatal • c = Number of incidents where layer was not present and fire was fatal • d = Number of incidents where layer was not present and fire was not fatal

Figure 4
Figure 4 breaks down the trend further into different types of dwelling, showing the average trends in the number of fires per year experienced by single occupancy houses, 1-3 storey flats, 4-9 storey flats, and 10+ storey flats.It is clear that the number of fires per year is decreasing across all of these property types.It seems from Figure 4A that the largest decrease in fires is in single occupancy houses; however, after normalizing by the total number of fires experienced by each property type in 2009 (Figure 4B), it seems that the

F
I G U R E 3 (A, left) Decrease in the number of all types of fires per million Londoners between 2009 and 2020.(B, right) Decrease in the number of dwelling fires per million dwellings between 2009 and 2020.F I G U R E 4 (A, left) Decrease in the number of dwelling fires each year across different property types.The total number of incidents in each category is given by the (n = …).(B, right) Decrease in the number of dwelling fires each year across different types of dwellings, normalized by the number of fires in that category in 2009.These trends show how much the frequency of fires has decreased, relative to how common fires were in that particular property type.

Figure 5
Figure5shows that about 83% of the injuries and 87% of the fatalities that occurred from 2009 to 2020 occurred in dwelling fires, despite the fact that this kind of fire made up only 28% of the fire incidents in the database.The higher severity (likelihood of a fire resulting in injuries and/or fatalities) of dwelling fires is a pattern commonly seen in fire incident data, with a study of home fires in the USA between 2014 and 201812 finding home structure fires to account for 28% of fires recorded but 77% of casualties.Fortunately, injuries and fatalities from these types of fires have decreased faster than other property categories in the UK, with 2020 having 44% fewer injuries

Figure 6
Figure 6 more clearly highlights the trends for injuries and fatalities in dwelling fires per million Londoners shown in Figure 5.It is clear that the number of injuries and fatalities due to dwelling fires decreased during the time period by a similar order of magnitude to the decrease in total number of dwelling fires per million Londoners (35%), with an average of 48 fewer dwelling injuries per million Londoners per year,

F I G U R E 9
Probability that once a fire has started it will become (A, left) injurious or (B, right) fatal across different property types, combining all dwelling fires between 2009 and 2020.Error bars give the 95% confidence interval.demonstrate that these kinds of fires are rare, meaning the average trends found in this report for the bulk of the LFB data may not reflect the trends for catastrophic fires, because these are outliers.One would instead need to either conduct an in-depth study of each catastrophic event to learn in detail what failed and what were the causes, as has happened with the Grenfell Tower fire in 2017, 30 or find enough examples of comparable catastrophic fires that they could be examined using statistical methods.The 3 catastrophic fires in London between 2009 and 2020 were the Lakanal House fire in 2009, 31 a fire in a single occupancy house in

F I G U R E 1 0
Frequency distribution of the number of injuries (A, left) and fatalities (B, right) in dwelling fires over the period of 2009-2020.Two of the three catastrophic fires are indicated by their name.This suggests that either London has a much lower percentage of households with working smoke alarms than the rest of the UK, or that either the survey results or the incident data are inaccurate.This could be investigated further in future work.Figure 11 also shows that the number of dwellings with at least one alarm present has increased by approximately 3% per year from 2009 to 2020, suggesting that the detection layer of protection has been improving in dwellings in London over this period.Predictably, the average evacuation time increases with the average storey height of the different dwelling types (based on a simple ordinary least-squares fit, using the median storey height for each category, average evacuation time increases by between 24 and 56 s per storey).It is worth noting that fire safety design in the UK generally assumes that occupants are able to reach a relative place of safety within 150 s (2.5 min) of a fire being detected.Although the average evacuation times here are generally longer than this, it is unclear exactly how firefighters define this evacuation time, with the definition in the IRS manual given as "Estimated time for completion of evacuation."Although not shown, average evacuation times remained relatively constant from 2009 to 2020.
if occupants are unable to evacuate the flat of fire origin, then it is likely that the fire would be fatal before the fire brigade arrived, and it may not be recorded as an evacuation having taken place.The details of the impact of these different protection layers should be investigated further in future work.F I G U R E 1 2 Changes in the proportions of dwelling fires where automatic sprinkler systems were present across different property types from 2009 to 2020.F I G U R E 1 3 Estimated increase in the odds of a fire being fatal based on failures in the compartmentation, detection, and evacuation layers of fire protection.Calculated using Equation (2).
would also be interesting to combine London Fire Brigade data with other datasets to gain further insight into of the observed trends.Such future studies would allow investigation of the impact, for example, of the age of buildings, the proportion of different types of buildings in different cities, the socioeconomic factors, and the materials used on the façade, to name a few.While the data used for the analysis presented here cannot comment on the specific factors of building age, construction materials, or whether particular social groups are at risk, it is in a sense considering the combined effects of all of these factors together.Though studying individual fires helps us learn how to improve specific aspects of building safety, this analysis shows us that the concept of layers of protection implemented in London is increasingly successful in reducing fire risk, and that Londoners should in general feel safe.However, this kind of analysis cannot be used to investigate catastrophic fires, which are rare and may be caused by many different factors, and so need to be investigated on an individual basis.It is therefore important for the fire safety profession to stay vigilant for this kind of fires, without forgetting that on average fire safety has improved.Also revealed is the incredible value of the London Fire Brigade, who have not only played a major role in contributing to the reduction in fires, fatalities, and injuries but have also gone to great lengths to collect these data that help them to continue improving safety in the future.F I G U R E 1 4 Illustrating the change in the total number of injuries and fatalities due to fires in London from 2009 to 2020.
Information on the evacuation, compartmentation, and suppression layers of fire protection during dwelling fires in different property types from 2009 to 2020.

4 |
CONCLUSIONSThe results of this analysis show that the overall fire risk for Londoners in general is the lowest since 2009 and is steadily reducing.In this time period, the number of fires each year has decreased despite the constant increase in the number of buildings and people in

TABLE 2 .
2escriptions of the features used for the analyses in this report, with the feature names given as found in the LFB incident database and the descriptions informed by the documentation of the national Incident Reporting System (IRS).2