Bantay Bangga

Data Collection

Data Exploration

Weekly Trends

Research Question 1: To what extent does hospital capacity affect the fatality rate of road incidents?

Research Question 2: What factors contribute to a hospital's overall capacity to provide care to patients?

As seen in the previous section, there are several factors that contribute to a hospital's overall capacity to provide care to its patients.

Modelling with Ordinary Least-Squares Linear Regression

From the data, certain linear trends are apparent. To supplement the findings above, an ordinary least-squares regression model (OLS) is trained to predict the number of injuries when given a subset of hospital capacity features (e.g., bed occupancy, medical staff readiness, and medical equipment availability). As an application of the project's findings, this model is in line with the research questions raised earlier.

Predicting the Minor Injuries

The first OLS model attempts to predict the total weekly number of minor injuries given the nationwide hospital capacity for that week. The baseline (null) model is one that always predicts the mean of the entire weekly aggregation, which is $\stackrel{-}{x}\approx 2.9767$. Hence, the null model has a root mean-squared error (RMSE) of $3.8849$ (i.e., the "score" to beat as a baseline).

In accordance with the findings in the previous section, the selected features for hospital capacity are:

1. the weekly total number of ICU beds occupied by COVID-19 patients (icu_o_c);
2. the weekly total number of vacant beds dedicated to COVID-19 patients (total_covid_v);
3. and the weekly total number of used mechanical ventilation units (mechvents_used).

It turns out that this subset of features yields the least RMSE out of the other subsets that have been exhaustively tried. A 10-fold cross-validation of the model reports $3.4178$ as the RMSE and $-8.4825$ as the $R^2$ score. Note that the RMSE of the OLS model is better than that of the null model. Unfortunately, the negative $R^2$ score is an indicator for a poorly fitted model, but this is the best that can be done with the data.

Predicting the Serious Injuries

The second OLS model attempts to predict the total weekly number of serious injuries given the nationwide hospital capacity for that week. The baseline (null) model uses the mean $\stackrel{-}{x}\approx 2.9767$. Hence, the null model has a root mean-squared error (RMSE) of $1.4058$ (i.e., the "score" to beat as a baseline).

In accordance with the findings in the previous section, the selected features for hospital capacity are:

1. the weekly total number of ICU beds occupied by non-COVID-19 patients (icu_o_nc);
2. the weekly total number of available beds (both ICU and non-ICU) (total_beds_v);
3. the weekly total number of vacant mechanical ventilation units (total_mechvent_v);
4. and the weekly total number of available surgical masks (surgmask).

A 10-fold cross-validation of the model reports an RMSE of $0.9627$ and an $R^2$ score of $-0.4616$ given this subset of features. An exhaustive search over all possible feature subsets shows that this yields the best metrics. Just like in the previous model, although the OLS model performs better than the null model, the slightly negative $R^2$ score indicates a poorly fitted model.

Predicting the Fatal Injuries

The final OLS model attempts to predict the total weekly number of fatal injuries given the nationwide hospital capacity for that week. The baseline (null) model uses the mean $\stackrel{-}{x}\approx 0.4651$. Hence, the null model has a root mean-squared error (RMSE) of $1.4996$ (i.e., the "score" to beat as a baseline).

In accordance with the findings in the previous section, the selected features for hospital capacity are:

1. the weekly total number of occupied beds (both ICU and non-ICU) (total_beds_o);
2. the weekly total number of used mechanical ventilation units (mechvents_used);
3. the weekly total number of available surgical masks (surgmask);
4. and the weekly total number of available N95 masks (n95).

A 10-fold cross-validation of the model reports an RMSE of $0.9143$ and an $R^2$ score of $-0.6221$ given this subset of features. An exhaustive search over all possible feature subsets shows that this yields the best metrics. Just like in the previous model, although the OLS model performs better than the null model, the slightly negative $R^2$ score indicates a poorly fitted model.

Simulations

Minor Injury Predictions

Given a total of:

• $600$ occupied ICU beds with COVID-19 patients
• $12400$ vacant beds dedicated for COVID-19 patients
• $0.42$ or $42\%$ mechanical ventilation unit usage rate

The number of predicted minor injuries given the model was $11.7770$.

Serious Injury Predictions

Given a total of:

• $3500$ occupied ICU beds with non-COVID-19 patients
• $48000$ total vacant beds (inclusive of both ICU and non-ICU beds)
• $3500$ vacant mechanical ventilation units
• $9500000$ available surgical masks

The number of predicted serious injuries given the model was $1.0188$.

Fatal Injury Predictions

Given a total of:

• $50000$ occupied beds (inclusive of ICU and non-ICU beds)
• $0.31$ or $31\%$ mechanical ventilation unit usage rate
• $10000000$ available surgical masks
• $1100000$ available N95 masks

The number of predicted serious injuries given the model was $0.6222$.

Conclusions

Features That Matter

The researchers have noticed that upon using machine learning and linear regression, as well as looking into correlations between each feature, that the occupancy of hospital beds can affect the injury rates of road-related incident victims across all injury types (minor, serious, fatal). The vacancy of hospital beds, which was intriguing as that would mean the amount of hospital beds, indifferent to whether or not they're occupied, can be attributed to different road injury types.

Similarly, some medical equipment also proved to be worth looking into. These were namely the two main types of masks used during the COVID-19 pandemic, which are surgical and N95 masks, and medical ventilation machines/units. These equipment may prove vital in reducing road-related incident fatalities in the near future. This might hint at the possibility of needing more equipment for the near future, especially when hospitals are tending to a certain sickness at the time.

Interestingly, no features under the medical staff dataset were deemed to be related to road-related incidents. For every injury type, bed occupancy and equipment availability proved to be more influencing than hospital staff availability. This may mean that the count of hospital staff is sufficient for aiding road victims.

Looking Forward

The absence of the medical staff in the predictors may signify a need to focus more on supplying hospital resources rather than hiring more people. This indicates that medical staff quantity, and availability in general, is not the main factor behind road-related injuries and fatalities. Hiring more people isn't the solution.

ICU bed occupancy is a major predictor worth considering and looking into, as it is present in all three injury types' list of predictors. This may hint at the importance of hospital decongestion, which may result in patients with certain illnesses and injuries gaining more priority over road victims. It may also be worth noting for hospitals and governments in distributing and allocating patients, should there be multiple hospitals in the vicinity.

Several features and data prediction techniques were not touched upon by the researchers. Further studies and work may be directed towards the effect of hospital proximity and distance from each other, as well as the distance from the road event to the hospital. More advanced models such as neural networks may also be used for a more in-depth and possibly more accurate prediction. Some features and terms including "hospital capacity" could also be separated into multiple features and therefore defined better.

Notably, the researchers conclude that, based from the model, "hospital capacity" (for tending after road crash victims) does not significantly rely on the availability of hospital staff. Perhaps more data is required to determine if this were true or not, as it seems counter-intuitive that more room for patients would not need more staff tending to their needs.

Pilipinas in a Nutshell (PILIPINUTS 2023)

Road incidents in the Philippines are commonly reported. A fatality rate of 25% of incidents was recorded in November 2020, in addition to widespread cases of minor and serious injuries resulting from road incidents. To avoid unnecessary fatalities, hospital capacity and quality of care should be kept at a sufficient level for these road crash victims.

However, as the COVID-19 pandemic demanded the world's full attention, this unique situation poses an interesting inquiry into what could possibly affect road crash incident fatalities and their possible avenues for recovery in a time where something else (i.e., the pandemic) was the top priority.

Throughout the data analysis, the following features were of interest to the researchers: hospital bed occupancy (with COVID-19 and non-COVID-19 beds being noted), hospital staff availability, and medical equipment availability. A significance level of 95% was set as the threshold for deciding on whether to consider linear relationships between features of hospital capacity and road-related fatality rates.

Among these features, minor and fatal injuries are related to the hospital bed occupancy through a linear combination consisting of the total number of COVID-19 patients in both ICU beds and non-ICU beds, which possibly hints at the larger role of beds in hospital decongestion. Unfortunately, staff resources showed no linear relationship with any of the injury types and cannot be considered a predictor. Perhaps more data is required here. Meanwhile, medical equipment is linearly related with serious injuries through a linear combination of surgical masks, gloves, and face shields.

These aforementioned features are used in an OLS regression model to predict the weekly total number of minor, serious, and fatal injuries. A 10-fold cross-validation of the trained model confirms that it performs better than the null model (which just predicts the mean at all times). However, the coefficient of determination is mostly negative, which hints at a possibly poorly fit model. Further work can expound on the feature set by considering hospital proximity to the road crash incident or use an outright different machine learning model altogether (i.e., neural networks).