CEMA/first_stage_script.R at main · DRIaDD/CEMA · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
# 0. load needed libraries
library(caret)
library(rnaturalearth)
library(readxl)
library(spdep)
library(mgcv)
library(broom)
library(tidyverse)
library(glm)
library(matrixStats)
filepath <- "c:/users/giselara/Documents/workshops/cema/"

clsi_factors <- read.csv(paste0(filepath, "guideline_factor.csv"))
clsi_factors <- clsi_factors[clsi_factors$CLSI_2023_adjustment_factor<1,]
tertiary_factors <- read.csv(paste0(filepath, "tertiary_factor.csv")) %>%
  mutate(tert_group = paste0(tert_group," ", super_region))
tertiary_factors$super_region <- NULL

raw_data <- read.csv("C:/Users/giselara/Documents/workshops/cema/raw_data.csv") %>%
  mutate(resistant = ifelse(interpretation %in% c('resistant'), 1, 0)) %>%
  group_by(country,super_region_name,year,guideline_group,tert_group,hosp_type,
           combo,combo2,species,antibiotic,abx_class) %>%
  summarise(total_test = n(), resistant = sum(resistant)) %>%
  mutate(prop = resistant / total_test)

###########################
# CLSI adjustment
############################

raw_data <- raw_data %>% left_join(clsi_factors)
raw_data$CLSI_2023_adjustment_factor[is.na(raw_data$CLSI_2023_adjustment_factor)] = 1

raw_data$resistant_adjusted <- raw_data$resistant * raw_data$CLSI_2023_adjustment_factor
raw_data <- raw_data %>% group_by(country, super_region_name, year,
                                  tert_group, hosp_type, combo,combo2,
                                  species, abx_class) %>%
  summarise(total_test = sum(total_test), resistant = sum(resistant_adjusted))


###########################
# Tertiary adjustments
############################
tertiary <- raw_data[(raw_data$hosp_type == "tertiary") & (raw_data$resistant > 0),] %>% left_join(tertiary_factors)
raw_data <- raw_data[(raw_data$hosp_type == "non-tertiary") | (raw_data$resistant == 0),]

tertiary$adj[is.na(tertiary$adj)] = 0
#raw_data$adj[raw_data$resistant <= 0] = 0
tertiary$prop <- tertiary$resistant / tertiary$total_test
tertiary$prop_adjusted <- exp(log(tertiary$prop) - tertiary$adj)
tertiary$resistant_adjusted <- tertiary$total_test * tertiary$prop_adjusted

summary(tertiary[,c('prop','prop_adjusted')])
tertiary$prop <- tertiary$prop_adjusted
tertiary$resistant <- tertiary$resistant_adjusted

tertiary[,c('resistant_adjusted','prop_adjusted','prop','adj','beta')] <- NULL

raw_data <- rbind(raw_data, tertiary)

raw_data <- raw_data %>% group_by(country, year, combo2) %>%
  summarise(total = sum(total_test), resistant = sum(resistant)) %>%
  mutate(prop = resistant / total)

##########################
# Linear model and prediction
##########################

abxcovs <- read.csv(paste0(filepath, 'abxcovs.csv'))# %>% filter(!is.na(location_name)) #%>% mutate(Country = location_name) %>% select(!location_name)
raw_data <- raw_data %>% left_join(abxcovs, by = c(c('country' = 'location_name'), c('year' = 'year_id')))

###########################
# Linear model for one combo and one country
############################
i='acinetobacter baumannii aminoglycoside'
combo_data <- raw_data[raw_data$combo2 == paste0(i) & raw_data$country == "United States of America",]

model1 <- lm(prop ~ year , data = combo_data)

print(model1)

combo_data$linear_full_pred <- predict(model1, type = 'response', allow.new.levels = TRUE)

print(cor(combo_data$prop,combo_data$linear_full_pred))

###########################
# Loop for linear and other modelling frameworks
############################

metric_table <- tibble()
table_cor <- c()
for(i in unique(raw_data$combo2)) {
  combo_data <- raw_data[raw_data$combo2 == paste0(i),]

  model1 <- lm(prop ~ year , data = combo_data)

  combo_data$linear_full_pred <- predict(model1, type = 'response', allow.new.levels = TRUE)

  response = cbind(successes = combo_data$resistant,
                   failures = (combo_data$total - combo_data$resistant))

  xg_grid <- expand.grid(nrounds = c(50, 100, 200),
                         max_depth = c(4, 6, 8),
                         eta = (3:6) / 100, #learning rate
                         colsample_bytree = .5,
                         min_child_weight = 1,
                         subsample = 1,
                         gamma = 0)

  xg_fit <- caret::train( prop ~ year + ddd_per_1000_fitted ,
                          data = combo_data,
                          #                       trControl = train_control,
                          verbose = T,
                          tuneGrid = xg_grid,
                          metric = "RMSE",
                          method = "xgbTree",
                          objective = "reg:logistic",iteration_range=200,
                          #trControl = cctrl,
                          # nrounds = 200000)
                          #n_estimators=1000, learning_rate=0.1, subsample=0.8, colsample_bytree=0.8,
                          #weights = combo_data$w
  )
  combo_data$xgboost_full_pred <- predict(xg_fit, combo_data, iteration_range=(best_iteration))

  ##########GAM

  covs_to_include = c('year','ddd_per_1000_fitted')
  gam_formula <- paste0('response ~ 1+ s(', paste(covs_to_include, collapse = ", bs = 'ts', k = 3) + s("), ", bs = 'ts', k = 3)")

  full_gam = mgcv::gam(response ~ 1 + s(year, bs = 'ts', k = 3),
                       data = combo_data,
                       family = 'quasibinomial',
                       #                     weights = mydata$w,
                       control = list(nthreads = 2))
  full_gam$model_name = 'GAM'

  combo_data$gam_full_pred <- predict(full_gam, combo_data, type = 'response')

  weight_table <-tibble()
  for(j in c("linear_full_pred"      ,    "gam_full_pred"  ,    "xgboost_full_pred" )) {
    mt <- c(model  = j )
    mt$cor <- cor(combo_data$prop,combo_data[,j])
    mt$combo <- i
    metric_table <- rbind(metric_table,mt)
    weight_table <- rbind(weight_table, mt)
  }

  # for(x in  1:length(child_models)){
  #   for(j in 1:length(child_models)){
  #     if(x==j){
  #     } else{
  #         cor1 <- c(child_models[x],child_models[j],formatC(cor(combo_data[,(paste0(child_models[x], '_full_pred'))],combo_data[,(paste0(child_models[j], '_full_pred'))])^2, format = 'f', digits = 6))
  #                 row1 <- c(cor1,combo = paste0(i))
  #         table_cor <- rbind(table_cor,row1)
  #         rm(cor1, row1)
  #        }
  #       }
  #     }
  #   }
  denominator <- weight_table %>% summarise(cor = sum(cor))
  weights <- weight_table$cor / as.numeric(denominator)
  X = as.matrix(combo_data[c("linear_full_pred"      ,    "gam_full_pred"  ,    "xgboost_full_pred")])
  combo_data$stacked_full_pred <- rowWeightedMeans(X, w = weights)
  stacked_pred <- combo_data[c("country", "year", "combo2", "stacked_full_pred")]

}
write.csv(combo_data, paste0(filepath,'combos.csv'), row.names = F)
write.csv(stacked_pred, paste0(filepath,'stacked.csv'), row.names = F)
write.csv(metric_table, paste0(filepath,'metric_table.csv'), row.names = F)
write.csv(table_cor, paste0(filepath,'table_cor.csv'), row.names = F)