[framework update] multimodal draft

[zzachw]

No description provided.

[jhnwu3]

Can we make a bandpass filter optional here? According to jathurshan:

• IIIC data may not use explicitly a bandpass filter
• ECOG data may do raw signal or band stop filtering

[jhnwu3]

Just want to make sure TextFeaturizer, BioSignal Featurizer are not too heavy or restrictive.

And, I also added some suggestions table processing for CXR and notes using the MIMIC4. But, I think we probably need to figure out how we want to initialize the dataset, and the pathing for it, because CXR and Notes will have different related pathings. We can probably ask the user to just put everything in one directory, or we can add additional optional filepathing variables and check if None else: parse_{table}.

[jhnwu3]

I think this is mostly fine. I think maybe any option for the user to define their own AutoTokenizer and AutoModel outside might be less heavy, because sometimes people may want to throw in their own existing finetuned model

[jhnwu3]

Looks good to me

[jhnwu3]

Hey Zhenbang, did we forget about MIMIC Note, and CXR here?

[jhnwu3]

def parse_discharge(self, patients: Dict[str, Patient]) -> Dict[str, Patient]:
    table = "discharge" # hardcoded, again might need user to explicitly download files into same directory/.
    df = pd.read_csv(os.path.join(self.tables_dir[table], f"{table}.csv"), 
                     dtype={"subject_id": str, "hadm_id": str})
    df = df.dropna(subset=["subject_id", "hadm_id", "text", "charttime"])
    df = df.sort_values(["subject_id", "hadm_id"], ascending=True)
    group_df = df.groupby("subject_id")
    def discharge_unit(p_id, p_info):
        events = []
        for v_id, v_info in p_info.groupby("hadm_id"):
            for text in v_info["text"]:
                attr_dict = {"text" : text, 
                                  "vocabulary" : "text",
                                  "visit_id"=v_id,
                                  "patient_id"=p_id}
                event = Event(
                    attr_dict = attr_dict,
                    timestamp=strptime(v_info["charttime"].values[0])
                )
                events.append(event)
        return events
    group_df = group_df.parallel_apply(
        lambda x: discharge_unit(x.subject_id.unique()[0], x)
    )
    patients = self._add_events_to_patient_dict(patients, group_df)
    return patients

def parse_cxr(self, patients: Dict[str, Patient]) -> Dict[str, Patient]:
      table = "cxr"
      cxr_file = "mimic-cxr-2.0.0-metadata"
      # hardcoded, might need to explicitly just have a CXR path in init.
      df = pd.read_csv(os.path.join(self.tables_dir[table], f"{cxr_file}.csv"),
                         dtype={"subject_id": str, "hadm_id": str})
      
      # combine date and time to create timestamp 
      df = df.dropna(subset=["subject_id", "study_id", "dicom_id"])
      df.StudyDate = df.StudyDate.astype(str)
      df.StudyTime = df.StudyTime.astype(str)
      # process all the dates and times
      df['StudyDateTime'] = df.apply(lambda row: self.transform_study_datetime(str(row['StudyDate']), str(row['StudyTime'])), axis=1)
      df = df.sort_values(["subject_id", "study_id"], ascending=True)
      
      group_df = df.groupby("subject_id")
      
      def cxr_unit(p_id, p_info):
          events = []
          for v_id, v_info in p_info.groupby("study_id"):
              for dicom_id, timestamp in zip(v_info["dicom_id"], v_info["StudyDateTime"]):
                  attr_dict = { "dicom_id"=dicom_id, # used for the dicom_id pathing
                      "vocabulary"="cxr"}
                  event = Event(
                      visit_id=v_id,
                      patient_id=p_id,
                      timestamp=strptime(timestamp)
                  )
                  events.append(event)
          return events
      
      group_df = group_df.parallel_apply(lambda x: cxr_unit(x.subject_id.unique()[0], x))
      patients = self._add_events_to_patient_dict(patients, group_df)
      return patients