Final

teams/elementsSubversius/README.md

@@ -0,0 +1,19 @@
+<img src="../../images/hachandhealth_banner.png" alt="" style="width:;"/>
+
+## EQUIPO: Elements Subversius
+
+Formado por:
+
+- Beatriz Martin
+- Juan Pablo López
+- Helena Solà
+- Imanol Morata
+
+Creación de modelo de clasificación para predecir si se va a producir un nuevo episodio de Bulimia o Anorexia a partir de los datos facilitados de la aplicación HealthApp.
+
+Hemos utilizado código R para preparar el dataset y Python para generar el modelo, usando Bayesian Ridge Regression para asignar una puntuación, y después buscando un límite que separe el episodio de la ausencia de episodio.
+
+
+---
+![](../../images/banner_es.png) 
+![](../../images/colaboradores_hack-health.png) 

teams/elementsSubversius/etc/README.md

@@ -0,0 +1,7 @@
+<img src="../../images/hachandhealth_banner.png" alt="" style="width:;"/>
+
+Poned en esta carpeta cualquier otro archivo que necesitéis subir.
+
+---
+![](../../images/banner_es.png) 
+![](../../images/colaboradores_hack-health.png) 

teams/elementsSubversius/slides/README.md

@@ -0,0 +1,7 @@
+<img src="../../images/hachandhealth_banner.png" alt="" style="width:;"/>
+
+Poned en esta carpeta las slides de vuestra presentación.
+
+---
+![](../../images/banner_es.png) 
+![](../../images/colaboradores_hack-health.png) 

teams/elementsSubversius/src/Python/FD_BR_classifier_test.py

@@ -0,0 +1,114 @@
+import numpy as np
+import pandas as pd
+import sys
+import matplotlib.pyplot as plt
+
+sys.path.insert(0,"C:/Braingaze/libraries")
+
+from sklearn import svm,neighbors,ensemble
+from sklearn.metrics import roc_curve, auc
+from sklearn.cross_validation import StratifiedKFold
+from sklearn.feature_selection import SelectKBest, f_classif
+from sklearn.linear_model import BayesianRidge
+
+import mc_module as mc
+
+### DATA PREPARATION ###
+
+df = pd.read_table('dtdt.csv',sep=',',encoding='mbcs').drop(['PATIENT','DAY'],axis=1).dropna()
+
+y = np.array(df['EP']) # Entry labelization
+np.place(y,y>0,1)
+
+ml = df.drop(['EP'],axis=1).copy()
+
+cols = ml.columns
+
+Z = np.array(ml).reshape(len(ml.index),len(cols))
+Z,y = np.vstack((Z[y==0],Z[y==1])), np.hstack((y[y==0],y[y==1]))
+len_ = np.min(np.floor(np.array([len(y[y==0])/2,len(y[y==1])/2,len(Z[0])])))
+
+### CROSS-VALIDATION AND GRID SEARCH ###
+
+classifier = BayesianRidge()
+cv = StratifiedKFold(y,n_folds=10)
+
+pars = []
+
+for k in np.arange(2,len_):
+
+    X = SelectKBest(f_classif,k=k).fit_transform(Z,y) # Select k best-separating features
+
+    for tsh in np.linspace(0.4,0.6,21):
+
+        acc = 0.
+
+        for i, (train, test) in enumerate(cv):
+            classifier.fit(X[train], y[train])
+            out = classifier.predict(X[test])
+            np.place(out,out>tsh,1) # Label according to different thresholds
+            np.place(out,out<=tsh,0)
+            acc += 1. - np.sum(np.mod(out + y[test],2))/len(y[test])
+
+        print(k,tsh,np.round(acc/(i+1),4)) # Accuracy for k features and threshold tsh
+        if len(pars) == 0:
+            pars = np.array([k,tsh,np.round(acc/(i+1),4)])
+        else:
+            pars = np.vstack((pars,np.array([k,tsh,np.round(acc/(i+1),4)])))
+
+### PLOT COLORFUL STUFF ###
+
+plt.scatter(pars[:,0],pars[:,1],c=pars[:,2],marker='+')
+plt.scatter(pars[np.argmax(pars[:,2])][0],pars[np.argmax(pars[:,2])][1],c='red',s=50,label='Est. accuracy: '+str(np.max(pars[:,2])))
+plt.legend()
+plt.xlabel('n features')
+plt.ylabel('threshold')
+plt.title('Bayesian Ridge regressor performance for FD episode prediction')
+plt.grid()
+plt.show()
+plt.clf()
+
+print(pars[np.argmax(pars[:,2])])
+
+k = pars[np.argmax(pars[:,2])][0]
+tsh = pars[np.argmax(pars[:,2])][1]
+
+kb = SelectKBest(f_classif,k=k)
+kb.fit(Z,y)
+X = kb.transform(Z)
+classifier.fit(X,y)
+
+t = classifier.predict(X)
+
+plt.scatter(y,t,c=y,s=50)
+plt.axhline(tsh,linestyle='--',color='k') # tsh may be changed to move along a ROC curve
+plt.grid()
+plt.xlabel('Class (0: no episode, 1: episode)')
+plt.ylabel('BR Score')
+plt.title('Class separation by Bayesian Ridge score')
+plt.show()
+plt.clf()
+
+### CHECK FOR OVER-FITTING ###
+
+U = np.random.uniform(0.,4.,size=np.shape(Z)) # Just create a random k-simplex and randomly label it
+t = np.floor(np.random.uniform(0.,2.,size=np.shape(y)))
+U = kb.transform(U) # Transform according to Skb rule
+
+classifier.fit(X,y)
+out = classifier.predict(U) # Predict random data
+
+# Plot for visual inspection of the randomized data
+
+plt.scatter(t,out,c=t,s=50)
+plt.axhline(tsh,linestyle='--',color='k')
+plt.grid()
+plt.xlabel('Class (0: no episode, 1: episode)')
+plt.ylabel('BR Score')
+plt.title('Random class separation')
+plt.show()
+plt.clf()
+
+np.place(out,out<=tsh,0)
+np.place(out,out>tsh,1)
+print(1. - np.sum(np.mod(out+t,2))/len(t)) # If not too far from 0.5 and FNR is similar fot FPR, everything fine