app.py

import os
HF_TOKEN = os.getenv("HF_TOKEN")

import numpy as np
import pandas as pd
import sklearn
import sklearn.metrics
from math import sqrt
from scipy import stats as st
from matplotlib import pyplot as plt

from sklearn.linear_model import LogisticRegression

import shap
import gradio as gr
import random
import re
import textwrap
from datasets import load_dataset


#Read data training data.

x1 = pd.read_csv("36m_data_train.csv", index_col = 0, low_memory = False)

x2 = pd.read_csv("60m_data_train.csv", index_col = 0, low_memory = False)

x3 = pd.read_csv("120m_data_train.csv", index_col = 0, low_memory = False)


#Read validation data.

x1_valid = pd.read_csv("36m_data_valid.csv", index_col = 0, low_memory = False)

x2_valid = pd.read_csv("60m_data_valid.csv", index_col = 0, low_memory = False)

x3_valid = pd.read_csv("120m_data_valid.csv", index_col = 0, low_memory = False)


#Define feature names.
f1_names = list(x1.columns)
f1_names = [f1.replace('__', ' - ') for f1 in f1_names]
f1_names = [f1.replace('_', ' ') for f1 in f1_names]

f2_names = list(x2.columns)
f2_names = [f2.replace('__', ' - ') for f2 in f2_names]
f2_names = [f2.replace('_', ' ') for f2 in f2_names]

f3_names = list(x3.columns)
f3_names = [f3.replace('__', ' - ') for f3 in f3_names]
f3_names = [f3.replace('_', ' ') for f3 in f3_names]


#Prepare training data for the outcome 1.
y1 = x1.pop('OUTCOME')

#Prepare validation data for the outcome 1.
y1_valid = x1_valid.pop('OUTCOME')

#Prepare training data for the outcome 2.
y2 = x2.pop('OUTCOME')

#Prepare validation data for the outcome 2.
y2_valid = x2_valid.pop('OUTCOME')

#Prepare training data for the outcome 3.
y3 = x3.pop('OUTCOME')

#Prepare validation data for the outcome 3.
y3_valid = x3_valid.pop('OUTCOME')


#Assign hyperparameters.

y1_params =  {'criterion': 'entropy', 'max_depth': 4, 'n_estimators': 1400, 'min_samples_leaf': 3, 'min_samples_split': 10, 'random_state': 31}
y2_params =  {'objective': 'binary', 'boosting_type': 'gbdt', 'lambda_l1': 0.6726024444744665, 'lambda_l2': 3.946412314107168e-08, 'num_leaves': 180, 'feature_fraction': 0.6838613576310666, 'bagging_fraction': 0.43284935253254003, 'bagging_freq': 5, 'min_child_samples': 83, 'metric': 'binary_logloss', 'verbosity': -1, 'random_state': 31}
y3_params =  {'objective': 'Logloss', 'colsample_bylevel': 0.08960988134854374, 'depth': 5, 'boosting_type': 'Plain', 'bootstrap_type': 'Bernoulli', 'subsample': 0.15476628999955983, 'random_seed': 31}


#Training models.

from sklearn.ensemble import RandomForestClassifier
rf = RandomForestClassifier(**y1_params)
y1_model = rf

y1_model = y1_model.fit(x1, y1)
y1_explainer = shap.Explainer(y1_model.predict, x1)
y1_calib_probs = y1_model.predict_proba(x1_valid)
y1_calib_model = LogisticRegression()
y1_calib_model = y1_calib_model.fit(y1_calib_probs, y1_valid)


from lightgbm import LGBMClassifier
lgb = LGBMClassifier(**y2_params)
y2_model = lgb

y2_model = y2_model.fit(x2, y2)
y2_explainer = shap.Explainer(y2_model.predict, x2)
y2_calib_probs = y2_model.predict_proba(x2_valid)
y2_calib_model = LogisticRegression()
y2_calib_model = y2_calib_model.fit(y2_calib_probs, y2_valid)


from catboost import CatBoostClassifier
cb = CatBoostClassifier(**y3_params)
y3_model = cb

y3_model = y3_model.fit(x3, y3)
y3_explainer = shap.Explainer(y3_model.predict, x3)
y3_calib_probs = y3_model.predict_proba(x3_valid)
y3_calib_model = LogisticRegression()
y3_calib_model = y3_calib_model.fit(y3_calib_probs, y3_valid)


output_y1 = (
    """          
        <br/>
        <center>The probability of 3-year survival:</center>
        <br/>
        <center><h1>{:.2f}%</h1></center>
"""
)

output_y2 = (
    """          
        <br/>        
        <center>The probability of 5-year survival:</center>
        <br/>        
        <center><h1>{:.2f}%</h1></center>
"""
)

output_y3 = (
    """          
        <br/>        
        <center>The probability of 10-year survival:</center>
        <br/>        
        <center><h1>{:.2f}%</h1></center>
"""
)


#Define predict for y1.
def y1_predict(*args):
    df1 = pd.DataFrame([args], columns=x1.columns)
    pos_pred = y1_model.predict_proba(df1)
    pos_pred = y1_calib_model.predict_proba(pos_pred)
    prob = pos_pred[0][1]
    prob = 1-prob
    output = output_y1.format(prob * 100)
    return output

#Define predict for y2.
def y2_predict(*args):
    df2 = pd.DataFrame([args], columns=x2.columns)
    pos_pred = y2_model.predict_proba(df2)
    pos_pred = y2_calib_model.predict_proba(pos_pred)        
    prob = pos_pred[0][1]
    prob = 1-prob
    output = output_y2.format(prob * 100)
    return output

#Define predict for y3.
def y3_predict(*args):
    df3 = pd.DataFrame([args], columns=x3.columns)
    pos_pred = y3_model.predict_proba(df3)
    pos_pred = y3_calib_model.predict_proba(pos_pred)            
    prob = pos_pred[0][1]
    prob = 1-prob    
    output = output_y3.format(prob * 100)
    return output


#Define function for wrapping feature labels.
def wrap_labels(ax, width, break_long_words=False):
    labels = []
    for label in ax.get_yticklabels():
        text = label.get_text()
        labels.append(textwrap.fill(text, width=width, break_long_words=break_long_words))
    ax.set_yticklabels(labels, rotation=0)
    

#Define interpret for y1.
def y1_interpret(*args):
    df1 = pd.DataFrame([args], columns=x1.columns)
    shap_values1 = y1_explainer(df1).values
    shap_values1 = np.abs(shap_values1)
    shap.bar_plot(shap_values1[0], max_display = 10, show = False, feature_names = f1_names)
    fig = plt.gcf()
    ax = plt.gca()
    wrap_labels(ax, 20)
    ax.figure
    plt.tight_layout()
    fig.set_figheight(7)
    fig.set_figwidth(9)
    plt.xlabel("SHAP value (impact on model output)", fontsize =12, fontweight = 'heavy', labelpad = 8)
    plt.tick_params(axis="y",direction="out", labelsize = 12)
    plt.tick_params(axis="x",direction="out", labelsize = 12)
    return fig

#Define interpret for y2.
def y2_interpret(*args):
    df2 = pd.DataFrame([args], columns=x2.columns)
    shap_values2 = y2_explainer(df2).values
    shap_values2 = np.abs(shap_values2)
    shap.bar_plot(shap_values2[0], max_display = 10, show = False, feature_names = f2_names)
    fig = plt.gcf()
    ax = plt.gca()
    wrap_labels(ax, 20)
    ax.figure
    plt.tight_layout()
    fig.set_figheight(7)
    fig.set_figwidth(9)
    plt.xlabel("SHAP value (impact on model output)", fontsize =12, fontweight = 'heavy', labelpad = 8)
    plt.tick_params(axis="y",direction="out", labelsize = 12)
    plt.tick_params(axis="x",direction="out", labelsize = 12)
    return fig

#Define interpret for y3.
def y3_interpret(*args):
    df3 = pd.DataFrame([args], columns=x3.columns)
    shap_values3 = y3_explainer(df3).values
    shap_values3 = np.abs(shap_values3)
    shap.bar_plot(shap_values3[0], max_display = 10, show = False, feature_names = f3_names)
    fig = plt.gcf()
    ax = plt.gca()
    wrap_labels(ax, 20)
    ax.figure
    plt.tight_layout()
    fig.set_figheight(7)
    fig.set_figwidth(9)
    plt.xlabel("SHAP value (impact on model output)", fontsize =12, fontweight = 'heavy', labelpad = 8)
    plt.tick_params(axis="y",direction="out", labelsize = 12)
    plt.tick_params(axis="x",direction="out", labelsize = 12)
    return fig



with gr.Blocks(title = "NCDB-Chordoma") as demo:
        
    gr.Markdown(
        """
    <br/>
    <center><h1>Spinal Chordoma</h1></center>
    <center><h2>Survival Outcomes Prediction Tool</h2></center>
    <center><i>The publication describing the details of this predictive tool will be posted here upon the acceptance of publication.</i><center>
        """
    )

    gr.Markdown(
        """
        <center><h3>Model Performances</h3></center>
          <div style="text-align:center;">
          <table style="width:100%;">
          <tr>
            <th>Outcome</th>
            <th>Algorithm</th>
            <th>Sensitivity</th>
            <th>Specificity</th>
            <th>Accuracy</th>
            <th>AUPRC</th>
            <th>AUROC</th>
            <th>Brier Score</th>
          </tr>
          <tr>
            <td>3-Year Mortality</td>
            <td>Random Forest</td>
            <td>0.321 (0.268 - 0.374)</td>
            <td>0.792 (0.746 - 0.838)</td>
            <td>0.708 (0.656 - 0.760)</td>
            <td>0.272 (0.221 - 0.323)</td>
            <td>0.742 (0.620 - 0.762)</td>
            <td>0.142 (0.102 - 0.182)</td>             
          </tr>
          <tr>
            <td>5-Year Mortality</td>
            <td>LightGBM</td>
            <td>0.590 (0.528 - 0.652)</td>
            <td>0.778 (0.726 - 0.83)</td>
            <td>0.714 (0.657 - 0.771)</td>
            <td>0.565 (0.502 - 0.628)</td>
            <td>0.807 (0.683 - 0.813)</td>
            <td>0.190 (0.14 - 0.24)</td>             
          </tr>
          <tr>
            <td>10-Year Mortality</td>
            <td>CatBoost</td>
            <td>0.862 (0.809 - 0.915)</td>
            <td>0.591 (0.515 - 0.667)</td>
            <td>0.788 (0.725 - 0.851)</td>
            <td>0.944 (0.908 - 0.980)</td>
            <td>0.893 (0.795 - 0.914)</td>
            <td>0.147 (0.092 - 0.202)</td>             
          </tr>                               
        </table>
        </div>
        """
    )    

    with gr.Row():

        with gr.Column():

            Age_at_Diagnosis = gr.Slider(label="Age", minimum = 18, maximum = 99, step = 1, value = 50)

            Sex = gr.Dropdown(label = "Sex", choices = ['Male', 'Female'], type = 'index', value = 'Male')
            
            Race = gr.Dropdown(label = "Race", choices = ['White', 'Black', 'Asian Indian or Pakistani', 'American Indian, Aleutian, or Eskimo', 'Chinese', 'Filipino', 'Vietnamese', 'Hawaiian', 'Japanese', 'Korean', 'Other or Unknown'], type = 'index', value = 'White')

            Hispanic_Ethnicity = gr.Dropdown(label = "Hispanic Ethnicity", choices = ['No', 'Yes', 'Unknown'], type = 'index', value = 'No')
            
            Insurance_Status = gr.Dropdown(label = "Insurance Status", choices = ['Private insurance', 'Medicare', 'Medicaid', 'Other government', 'Not insured', 'Unknown'], type = 'index', value = 'Private insurance')
            
            Facility_Type = gr.Dropdown(label = "Facility Type", choices = ['Academic/Research Program', 'Comprehensive Community Cancer Program', 'Integrated Network Cancer Program', 'Community Cancer Program', 'Other or Unknown'], type = 'index', value = 'Academic/Research Program')
            
            Facility_Location = gr.Dropdown(label = "Facility Location", choices = ['South Atlantic', 'East North Central', 'Middle Atlantic', 'East North Central', 'Middle Atlantic', 'Pacific', 'West South Central', 'West North Central', 'East South Central', 'New England', 'Mountain', 'Unknown or Other'], type = 'index', value = 'South Atlantic')

            CharlsonDeyo_Score = gr.Dropdown(label = "Charlson-Deyo Score", choices = ['0', '1', '2', 'Greater than 3'], type = 'index', value = '0')
                                    
            Histology = gr.Dropdown(label = "Histology", choices = ['Chordoma, NOS', 'Chondroid Chordoma', 'Dedifferentiated Chordoma'], type = 'index', value = 'Chordoma, NOS')
            
            Primary_Site = gr.Dropdown(label = "Primary Site", choices = ['Sacrum/Pelvis', 'Spine'], type = 'index', value = 'Sacrum/Pelvis')
            
            Diagnostic_Biopsy = gr.Dropdown(label = "Diagnostic Biopsy", choices = ['No', 'Yes', 'Unknown'], type = 'index', value = 'No')
            
            Tumor_Size_Largest_Diameter = gr.Dropdown(label = "Tumor Size (Largest Diameter)", choices = ['< 2 cm', '2 - 3.9 cm', '4 - 5.9 cm', '6 - 7.9 cm', '8 - 9.9 cm', '10 - 11.9 cm', '12 - 13.9 cm', '14 - 15.9 cm', '16 - 17.9 cm', '18 - 19.9 cm', '> 20 cm', 'Unknown'], type = 'index', value = '< 2 cm')
            
            Tumor_Size_Second_Largest_Diameter = gr.Dropdown(label = "Tumor Size (Second Largest Diameter)", choices = ['< 2 cm', '2 - 3.9 cm', '4 - 5.9 cm', '6 - 7.9 cm', '8 - 9.9 cm', '10 - 11.9 cm', '12 - 13.9 cm', '14 - 15.9 cm', '16 - 17.9 cm', '18 - 19.9 cm', '> 20 cm', 'Unknown'], type = 'index', value = '< 2 cm')
            
            Tumor_Size_Third_Largest_Diameter = gr.Dropdown(label = "Tumor Size (Third Largest Diameter)", choices = ['< 2 cm', '2 - 3.9 cm', '4 - 5.9 cm', '6 - 7.9 cm', '8 - 9.9 cm', '10 - 11.9 cm', '12 - 13.9 cm', '14 - 15.9 cm', '16 - 17.9 cm', '18 - 19.9 cm', '> 20 cm', 'Unknown'], type = 'index', value = '< 2 cm')            
                                    
            Regional_Lymph_Nodes = gr.Dropdown(label = 'Regional Lymph Nodes', choices = ['No', 'Yes', 'Unknown or not applicable'], type = 'index', value = 'No')
            
            Distant_Metastasis = gr.Dropdown(label = 'Distant Metastasis', choices = ['No', 'Yes', 'Unknown or not applicable'], type = 'index', value = 'No')
            
            Surgery = gr.Dropdown(label = "Surgery", choices = ['No', 'Yes', 'Unknown'], type = 'index', value = 'Yes')
            
            Surgical_Margins = gr.Dropdown(label = "Surgical Margins", choices = ['No residual tumor', 'Residual tumor', 'No surgery was performed', 'Unknown'], type = 'index', value = 'No residual tumor')
            
            Radiation_Treatment = gr.Dropdown(label = "Radiation Treatment", choices = ['No', 'Yes', 'Unknown'], type = 'index', value = 'Yes')
            
            Chemotherapy = gr.Dropdown(label = "Chemotherapy", choices = ['No', 'Yes', 'Unknown'], type = 'index', value = 'Yes')            
            
        with gr.Column():
            
            with gr.Box():
                
                gr.Markdown(
                    """
                    <center> <h2>3-Year Survival</h2> </center>
                    <br/>
                    <center> This model uses the Random Forest algorithm.</center>
                    <br/>
                    """
                    )
                
                with gr.Row():
                    y1_predict_btn = gr.Button(value="Predict")
                
                gr.Markdown(
                    """
                    <br/>
                    """
                    )
                
                label1 = gr.Markdown()
                
                gr.Markdown(
                    """
                    <br/>
                    """
                    )
                
                with gr.Row():
                    y1_interpret_btn = gr.Button(value="Explain")
                
                gr.Markdown(
                    """
                    <br/>
                    """
                    )
                
                plot1 = gr.Plot()
                
                gr.Markdown(
                    """
                    <br/>
                    """
                    )
                
            with gr.Box():
                gr.Markdown(
                    """
                    <center> <h2>5-Year Survival</h2> </center>
                    <br/>
                    <center> This model uses the LightGBM algorithm.</center>
                    <br/>
                    """
                    )
                
                with gr.Row():
                    y2_predict_btn = gr.Button(value="Predict")
                
                gr.Markdown(
                    """
                    <br/>
                    """
                    )
                
                label2 = gr.Markdown()
                
                gr.Markdown(
                    """
                    <br/>
                    """
                    )
                
                with gr.Row():
                    y2_interpret_btn = gr.Button(value="Explain")
                
                gr.Markdown(
                    """
                    <br/>
                    """
                    )
                
                plot2 = gr.Plot()
                
                gr.Markdown(
                    """
                    <br/>
                    """
                    )
                
            with gr.Box():
                
                gr.Markdown(
                    """
                    <center> <h2> 10-Year Survival</h2> </center>
                    <br/>
                    <center> This model uses the CatBoost algorithm.</center>
                    <br/>
                    """
                    )
                
                with gr.Row():
                    y3_predict_btn = gr.Button(value="Predict")
                
                gr.Markdown(
                    """
                    <br/>
                    """
                    )
                
                label3 = gr.Markdown()
                
                gr.Markdown(
                    """
                    <br/>
                    """
                    )
                
                with gr.Row():
                    y3_interpret_btn = gr.Button(value="Explain")
                
                gr.Markdown(
                    """
                    <br/>
                    """
                    )
                
                plot3 = gr.Plot()
                
                gr.Markdown(
                    """
                    <br/>
                    """
                    )  

           
                y1_predict_btn.click(
                    y1_predict,
                    inputs = [Facility_Type, Facility_Location, Age_at_Diagnosis, Sex, Race, Hispanic_Ethnicity, Insurance_Status, CharlsonDeyo_Score, Primary_Site, Histology, Diagnostic_Biopsy, Tumor_Size_Largest_Diameter, Tumor_Size_Second_Largest_Diameter, Tumor_Size_Third_Largest_Diameter, Regional_Lymph_Nodes, Distant_Metastasis, Surgery, Surgical_Margins, Chemotherapy, Radiation_Treatment],
                    outputs = [label1]
                )

                y2_predict_btn.click(
                    y2_predict,
                    inputs = [Facility_Type, Facility_Location, Age_at_Diagnosis, Sex, Race, Hispanic_Ethnicity, Insurance_Status, CharlsonDeyo_Score, Primary_Site, Histology, Diagnostic_Biopsy, Tumor_Size_Largest_Diameter, Tumor_Size_Second_Largest_Diameter, Tumor_Size_Third_Largest_Diameter, Regional_Lymph_Nodes, Distant_Metastasis, Surgery, Surgical_Margins, Chemotherapy, Radiation_Treatment],
                    outputs = [label2]
                )
                
                y3_predict_btn.click(
                    y3_predict,
                    inputs = [Facility_Type, Facility_Location, Age_at_Diagnosis, Sex, Race, Hispanic_Ethnicity, Insurance_Status, CharlsonDeyo_Score, Primary_Site, Histology, Diagnostic_Biopsy, Tumor_Size_Largest_Diameter, Tumor_Size_Second_Largest_Diameter, Tumor_Size_Third_Largest_Diameter, Regional_Lymph_Nodes, Distant_Metastasis, Surgery, Surgical_Margins, Chemotherapy, Radiation_Treatment],
                    outputs = [label3]
                )

                y1_interpret_btn.click(
                    y1_interpret,
                    inputs = [Facility_Type, Facility_Location, Age_at_Diagnosis, Sex, Race, Hispanic_Ethnicity, Insurance_Status, CharlsonDeyo_Score, Primary_Site, Histology, Diagnostic_Biopsy, Tumor_Size_Largest_Diameter, Tumor_Size_Second_Largest_Diameter, Tumor_Size_Third_Largest_Diameter, Regional_Lymph_Nodes, Distant_Metastasis, Surgery, Surgical_Margins, Chemotherapy, Radiation_Treatment],
                    outputs = [plot1],
                )
                
                y2_interpret_btn.click(
                    y2_interpret,
                    inputs = [Facility_Type, Facility_Location, Age_at_Diagnosis, Sex, Race, Hispanic_Ethnicity, Insurance_Status, CharlsonDeyo_Score, Primary_Site, Histology, Diagnostic_Biopsy, Tumor_Size_Largest_Diameter, Tumor_Size_Second_Largest_Diameter, Tumor_Size_Third_Largest_Diameter, Regional_Lymph_Nodes, Distant_Metastasis, Surgery, Surgical_Margins, Chemotherapy, Radiation_Treatment],
                    outputs = [plot2],
                )

                y3_interpret_btn.click(
                    y3_interpret,
                    inputs = [Facility_Type, Facility_Location, Age_at_Diagnosis, Sex, Race, Hispanic_Ethnicity, Insurance_Status, CharlsonDeyo_Score, Primary_Site, Histology, Diagnostic_Biopsy, Tumor_Size_Largest_Diameter, Tumor_Size_Second_Largest_Diameter, Tumor_Size_Third_Largest_Diameter, Regional_Lymph_Nodes, Distant_Metastasis, Surgery, Surgical_Margins, Chemotherapy, Radiation_Treatment],
                    outputs = [plot3],
                )
                
    gr.Markdown(
                """    
                <center><h2>Disclaimer</h2>
                <center> 
                The data utilized for this tool is sourced from the Commission on Cancer (CoC) of the American College of Surgeons and the American Cancer Society. These institutions, however, have not verified the information and are not responsible for the statistical validity of the data analysis or the conclusions drawn by the authors. This predictive tool, available on this webpage, is designed to provide general health information only and is not a substitute for professional medical advice, diagnosis, or treatment. It is strongly recommended that users consult with their own healthcare provider for any health-related concerns or issues. The authors make no warranties or representations, express or implied, regarding the accuracy, timeliness, relevance, or utility of the information contained in this tool. The health information in the prediction tool is subject to change and can be affected by various confounders, therefore it may be outdated, incomplete, or incorrect. No doctor-patient relationship is created by using this prediction tool and the authors have not validated its content. The authors do not record any specific user information or initiate contact with users. Before making any healthcare decisions or taking or refraining from any action based on the information in this prediction tool, it is advisable to seek professional advice from a healthcare provider. By using the prediction tool, users acknowledge and agree that neither the authors nor any other party will be liable for any decisions made, actions taken or not taken as a result of the information provided herein.
                <br/>
                <h4>By using this tool, you accept all of the above terms.<h4/>
                </center>
                """
    )                
                
demo.launch()