app.py

import streamlit as st
import json
import random
import pandas as pd
import pickle
import json

# set page configuration to wide mode
st.set_page_config(layout="wide")

# section 1
st.markdown("#### About")
st.markdown("Differential diagnosis of dementia remains a challenge in neurology due to symptom overlap across etiologies, yet it is crucial for formulating early, personalized management strategies. Here, we present an AI model that harnesses a broad array of data, including demographics, individual and family medical history, medication use, neuropsychological assessments, functional evaluations, and multimodal neuroimaging, to identify the etiologies contributing to dementia in individuals.")
st.markdown("Links:\n* Paper: [https://www.nature.com/articles/s41591-024-03118-z](https://www.nature.com/articles/s41591-024-03118-z)\n* GitHub: [https://github.com/vkola-lab/nmed2024](https://github.com/vkola-lab/nmed2024)\n* Our lab: [https://vkola-lab.github.io/](https://vkola-lab.github.io/)")

# section 2
st.markdown("#### Demo")
st.markdown("This Hugging Face Space is published for demonstration purposes. Users can input over 300 clinical entries to assess the etiologies contributing to cognitive impairment. However, due to the computational power limitations of the Hugging Face free tier, imaging features and Shapley values analysis are not supported. For the full implementation, please refer to our GitHub repository.")
st.markdown("To use the demo:\n* Provide input features in the form below. Feature missing is allowed.\n* Click the \"**RANDOM EXAMPLE**\" button to populate the form with a randomly selected datapoint.\n* Use the \"**PREDICT**\" button to submit all input features for assessment, then the predictions will be posted in a table.")

# load model
@st.cache_resource
def load_model():
    import adrd
    try:   
        ckpt_path = './ckpt_swinunetr_stripped_MNI.pt'
        model = adrd.model.ADRDModel.from_ckpt(ckpt_path, device='cpu')
    except:
        # ckpt_path = '../adrd_tool_copied_from_sahana/dev/ckpt/ckpt_swinunetr_stripped_MNI.pt'
        ckpt_path = '/data_1/skowshik/ckpts_backbone_swinunet/ckpt_swinunetr_stripped_MNI.pt'
        model = adrd.model.ADRDModel.from_ckpt(ckpt_path, device='cpu')
    return model

model = load_model()

def predict_proba(data_dict):
    pred_dict = model.predict_proba([data_dict])[1][0]
    return pred_dict

# load meta data csv
file_path = './data/input_meta_info.csv'
input_meta_info = pd.read_csv(file_path)

# load NACC testing data
from data.dataset_csv import CSVDataset
dat_tst = CSVDataset(
    dat_file = "./data/test_public.csv", 
    cnf_file = "./data/input_meta_info.csv"
)

def get_random_example():
    idx = random.randint(0, len(dat_tst) - 1)
    random_case = dat_tst[idx][0]
    return random_case

# Get random example features if the button is clicked
if 'random_example' not in st.session_state:
    st.session_state.random_example = None

st.markdown('---')
cols = st.columns(3)
with cols[1]:
    random_example_button = st.button("RANDOM EXAMPLE", use_container_width=True)
if random_example_button:
    st.session_state.random_example = get_random_example()
    st.rerun()

random_example = st.session_state.random_example

def create_input(df, i):
    row = df.iloc[i]
    name = row['Name']
    description = row['Description']

    # dirty work, inspect keys and values
    values = row['Values']
    values = values.replace('\'', '\"')
    values = values.replace('\"0\": nan, ', '')
    values = json.loads(values)

    for k, v in list(values.items()):
        if v == 'Unknown':
            values.pop(k)
        elif k in ('9', '99', '999'):
            values.pop(k)

        # get default value from random example if available
    default_value = random_example[name] if random_example and name in random_example else None
    if type(default_value) is float:
        default_value = int(default_value)

    # Determine the type of widget based on values
    if 'range' in values:
        if ' - ' in values['range']:
            min_value, max_value = map(float, values['range'].split(' - '))
            min_value, max_value = int(min_value), int(max_value)

            if default_value is not None:
                if default_value > max_value or default_value < min_value:
                    default_value = None
                
            st.number_input(description, key=name, min_value=min_value, max_value=max_value, value=default_value, placeholder=values['range'])
        else:
            min_value = int(values['range'].replace('>= ', ''))
            if default_value is not None:
                if default_value < min_value or default_value == 8888:
                    default_value = None

            st.number_input(description, key=name, min_value=min_value, value=default_value, placeholder=values['range'])
    else:
        values = {int(k): v for k, v in values.items()}
        if default_value in values:
            default_index = list(values.keys()).index(default_value)
        else:
            default_index = None

        st.selectbox(
            description, 
            options = values.keys(), 
            key = name, 
            index = default_index,
            format_func=lambda x: values[x]
        )

# create form
with st.form("dynamic_form"):
    sections = input_meta_info['Section'].unique()
    for section in sections:

        with st.container():
            st.markdown(f"##### {section}")
            sub_df = input_meta_info[input_meta_info['Section'] == section]

            cols = st.columns(3)
            with cols[0]:
                for i in range(0, len(sub_df), 3):
                    create_input(sub_df, i)
            with cols[1]:
                for i in range(1, len(sub_df), 3):
                    create_input(sub_df, i)
            with cols[2]:
                for i in range(2, len(sub_df), 3):
                    create_input(sub_df, i)    

        # seperate line
        st.markdown("---")

    cols = st.columns(3)
    with cols[1]:
        predict_button = st.form_submit_button("PREDICT", use_container_width=True, type='primary')

# load mapping
with open('./data/nacc_variable_mappings.pkl', 'rb') as file:
    nacc_mapping = pickle.load(file)

def convert_dictionary(original_dict, mappings):
    transformed_dict = {}
    
    for key, value in original_dict.items():
        if key in mappings:
            new_key, transform_map = mappings[key]
            
            # If the value needs to be transformed
            if value in transform_map:
                transformed_value = transform_map[value]
            else:
                transformed_value = value  # Keep the original value if no transformation is needed
            
            transformed_dict[new_key] = transformed_value
    
    return transformed_dict

if predict_button:
    # get form input
    names = input_meta_info['Name'].tolist()
    data_dict = {}
    for name in names:
        data_dict[name] = st.session_state[name]
    
    # convert
    data_dict = convert_dictionary(data_dict, nacc_mapping)
    pred_dict = predict_proba(data_dict)

    # change key name and value representations
    key_mappings = {
        'NC': 'Normal cognition',
        'MCI': 'Mild cognitive impairment',
        'DE': 'Dementia',
        'AD': 'Alzheimer\'s disease',
        'LBD': 'Lewy bodies and Parkinson\'s disease',
        'VD': 'Vascular brain injury or vascular dementia including stroke',
        'PRD': 'Prion disease including Creutzfeldt-Jakob disease',
        'FTD': 'Frontotemporal lobar degeneration',
        'NPH': 'Normal pressure hydrocephalus',
        'SEF': 'Systemic and external factors',
        'PSY': 'Psychiatric diseases',
        'TBI': 'Traumatic brain injury',
        'ODE': 'Other causes which include neoplasms, multiple systems atrophy, essential tremor, Huntington\'s disease, Down syndrome, and seizures'
    }
    pred_dict = {key_mappings[k]: f"{v * 100:.2f}%" for k, v in pred_dict.items()}

    df = pd.DataFrame(list(pred_dict.items()), columns=['Label', 'Predicted probability'])
    st.table(df)