Ehr transformer ICU example

examples/fuse_examples/multimodality/ehr_transformer/__init__.py

@@ -0,0 +1 @@
+__version__ = "0.0.1"

examples/fuse_examples/multimodality/ehr_transformer/config_PhysioNetCinC.yaml

@@ -0,0 +1,130 @@
+dataset:
+  raw_data_path: 'C:/D_Drive/Projects/EHR_Transformer/PhysioNet/predicting-mortality-of-icu-patients-the-physionetcomputing-in-cardiology-challenge-2012-1.0.0/predicting-mortality-of-icu-patients-the-physionet-computing-in-cardiology-challenge-2012-1.0.0'
+  #raw_data_pickle: 'raw_data.pkl'
+  split_filename: None
+  num_folds: 5
+  train_folds: [ 0, 1, 2 ]
+  validation_folds: [ 3 ]
+  test_folds: [ 4 ]
+  seed: 2580
+  reset_split: True
+  num_percentiles : 4 #number of bins (percentiles) for converting floating lab/vital measurements to categorical values
+  categorical_max_num_of_values : 5 #max number of uniq values for categorical variable for not to be digitized
+  min_hours_in_hospital: 46
+  visit_to_embed_static_variables: FIRST #embed to first (FIRST) or to all visits (ALL) or null (not to embed)
+  static_variables_to_embed: ['Age','ICUType','Height','Weight','BMI',]
+
+data:
+  #### Test/Val set defs
+  days_to_inddate: 60  #90 #None #180 #only visits before days_to_ind to index date (or after index date, if negative) are considered for test/val
+  days_to_inddate_start: null #only the visits between ind_date-days_to_inddate_start and ind_date-days_to_inddate are considered for test/val
+  event_prediction_window_days: null  #90
+  treatment_event_prediction_window_days: null  #60
+  #### Train set defs
+  # in case of train we take all visits (we want to train on as much data as we can), so days_to_inddate_tr and days_to_inddate_start_tr should be both null
+  days_to_inddate_tr: 60  #90 #None #180 #only visits before days_to_ind to index date (or after index date, if negative) are considered for train
+  days_to_inddate_start_tr: null #only the visits between days_to_inddate_start and days_to_inddate are considered for train
+  event_prediction_window_days_tr: null  #90
+
+  visit_days_resolution: 10 #consecutive visits that are less than visit_days_resolution days apart are merged into one visit.
+    #Note that merging stops only when number of days between consecutive visits is more than visit_days_resolution, so
+    #within a group of merged visits there may be some more than visit_days_resolution days apart.
+    #E.g. for visit days 1, 3, 5, 9, and for visit_days_resolution=2, days 1, 3, and 5 will be merged into a single visit, and day 10 will be a separate visit
+
+  limit_visits: null #210000 #null #50000000 #10000 #None #don't take too few (e.g. less than 10k), otherwise you'll get only a single visit per patient, which is not enough in case of next visit prediction
+  out_type: EVENT #EVENT or LABEL - determines the type of outcome GT, whether it is a list of events, or a binary label
+  task: CKD   #RC #"RC24" #"PD"#"BC"#'EP' #The task is only used output path naming and in db access (ehrtransformers.data_access.db2_data), to choose which dataset to read
+  subtask: ALL_SERV
+  task_type: outcome
+  output_name: admdate_visit
+  use_procedures: false #true #false
+  num_loader_workers: 1
+  input_column_names: [CCS, ]
+  min_visits_per_patient: 10 #patients with less visits are ignored
+
+  # When predicting next visit, this will map regular visit dictionary (codes to indices, index per code) to reduced
+  # visit dictionary (codes to indices, code groups mapped to (fewer) indices). When no mapping is used, this is null.
+  reduce_pred_visit_vocab_mapping_path: null #/projects/msieve_dev3/usr/vadim/my_work/Code/BEHRT/ehrtransformers/data_access/icd_to_ccs.json #null
+
+
+  data_source: csv #sql_db #csv
+  data_source_str: '/home/simona/Dev/data/ckd5/ckd5_all_transformer.csv'      #/gpfs/haifa/projects/m/msieve3/usr/Capable/MS_renal/data/rc_for_ehr_final.csv
+  #categorial_codes_str: /gpfs/haifa/projects/m/msieve3/usr/Capable/MS_renal/data/categories_codes_v2.csv
+  gt_source_str: null
+  sample_data_source_str: '/home/simona/Dev/data/ckd5/ckd5_all_transformer_10000.csv'    #/data/usr/vadim/EHR/rc_for_ehr_sample.csv
+
+learning:
+  batch_size: 100 #64 300, 480, 420
+  device: cuda:0 #cpu
+  is_data_parallel: false
+  gradient_accumulation_steps: 1
+  save_model: true
+  stop_file: stop.txt
+  optimization:
+    lr: 1.0e-05 #1.0e-5 #3.0e-06 #1e-4, #3e-6,
+    warmup_proportion: 0.1
+    weight_decay: 0.01
+global:
+  age_symbol: null
+  debug_mode: true #if false, outputs to a log file instead of the main output
+  global_stat_file: global_stats.csv
+  max_age: 110
+  max_len_seq: 100 #How many visits in a BEHRT input vector
+  max_label_len: 10 #How many classes in the multiclass task TODO: remove, this is no longer used
+  min_visit: 5 #Visit trajectories with fewer visits are ignored
+  month: 1 # choice whether to use months for age (=1) or years (=12)
+  uber_base_path: '/home/simona/Dev/EHR_experiments/ckd5'   #/data/usr/simona/EHR/ #path utils to all EHR experiments
+model:
+  age_vocab_size: null # number of vocab for age embedding
+  attention_probs_dropout_prob: 0.22 # multi-head attention dropout rate
+  hidden_act: gelu # activation function ("gelu", 'relu', 'swish')
+  hidden_dropout_prob: 0.2 # dropout rate
+  hidden_size: 240 #48  #240 #72 #120 #288, # word embedding and seg embedding hidden size (needs to be a multiple of attention heads)
+  initializer_range: 0.02 # parameter weight initializer range
+  intermediate_size: 512 #128 #512 # the size of the "intermediate" layer in the transformer encoder
+  max_position_embedding: 100 # maximum number of tokens
+  num_attention_heads: 24 # number of attention heads
+  num_hidden_layers: 6 # number of multi-head attention layers required
+#  pretrain_model: /data/usr/vadim/EHR/PD_ALL_SERV_270_after_ind_90_to_event_tr_90_to_event_merge_5_days/EVENT/models/PD/outcome_admdate_visit/run_139/Logs/model_dir/checkpoint_last_epoch.pth # pretrained model path, None if not to continue from pretrained
+#  pretrain_model: /data/usr/vadim/EHR/PD_ALL_SERV_270_after_ind_90_to_event_tr_90_to_event_merge_5_days/EVENT/models/PD/outcome_admdate_visit/run_136/Logs/model_dir/checkpoint_last_epoch.pth # pretrained model path, None if not to continue from pretrained
+#  pretrain_model: /data/usr/vadim/EHR/PD_ALL_SERV_270_after_ind_90_to_event_tr_90_to_event_merge_10_days/EVENT/models/PD/outcome_admdate_visit/run_104/Logs/model_dir/checkpoint_last_epoch.pth # pretrained model path, None if not to continue from pretrained
+#  pretrain_model: /data/usr/vadim/EHR/PD_ALL_SERV_270_after_ind_90_to_event_tr_90_to_event_merge_10_days_with_procedures/EVENT/models/PD/outcome_admdate_visit/run_113/Logs/model_dir/checkpoint_last_epoch.pth # pretrained model path, None if not to continue from pretrained
+#  pretrain_model: /data/usr/vadim/EHR/PD_ALL_SERV_270_after_ind_90_to_event_tr_90_to_event_merge_5_days_with_procedures/EVENT/models/PD/outcome_admdate_visit/run_105/Logs/model_dir/checkpoint_last_epoch.pth
+#  pretrain_model: /data/usr/vadim/EHR/RC_ALL_SERV_90_to_ind_tr_90_to_ind_90_to_event_tr_90_to_event_merge_10_days/EVENT/models/RC/outcome_admdate_visit/run_129/Logs/model_dir/checkpoint_last_epoch.pth
+#  pretrain_model: /home/simona/Dev/EHR_experiments/ckd5/CKD_ALL_SERV_60_to_ind_tr_60_to_ind_merge_10_days/EVENT/models/CKD/outcome_admdate_visit/run_101/Logs/model_dir/checkpoint_last_epoch.pth
+#  pretrain_model: /home/simona/Dev/EHR_experiments/ckd5/CKD_ALL_SERV_60_to_ind_tr_60_to_ind_merge_10_days/EVENT/models/CKD/outcome_admdate_visit/run_177/Logs/model_dir/last.ckpt
+  pretrain_model: null
+  seg_vocab_size: null # number of vocab for seg embedding
+  train_epochs: 2
+  vocab_size: null # number of disease + symbols for word embedding
+  reverse_input_direction: true
+  heads: [gender, next_vis, disease_prediction] #, response_profile, gender] #[event, next_vis, gender, cluster, contrastive, response_profile]
+  head_weights: [0.005, 32.0, 0.02]
+  sampler_weights: [0.9, 0.1]
+  contrastive_instances: 1 #number of patient visits used in contrastive loss (reducing distance between embeddings of the same patient)
+
+naming_conventions:
+  age_key: AGE
+  age_month_key: AGE_MON
+  date_key: ADMDATE
+  index_date_key: INDDATE
+  svc_date_key: SVCDATE
+  adm_date_key: ADMDATE
+  date_birth_key: DOBYR
+  dxver_key: DXVER
+  diagnosis_vec_key: DX
+  gender_key: SEX
+  male_val: 1 #'1'
+  female_val: 2 #'2'
+  label_key: label
+  outcome_key: CKD  #RC #Disease column
+  event_key: EVENTS #event column
+  treatment_event_key: TREATMENT_EVENTS
+  patient_id_key: ENROLID
+  disease_key: CKD  #RC #this is not used
+  healthy_val: 0
+  sick_val: 1
+  split_key: split
+  fold_key: fold
+  next_visit_key: DXNEXTVIS
+