Forecas(ng Poten(al Diabetes Complica(ons

advertisement
The Twenty-­‐Eighth AAAI Conference on ArtiAicial Intelligence (AAAI-­‐14, Quebec, Canada) Paper ID: 534 Forecas(ng Poten(al Diabetes Complica(ons Yang Yang, Walter Luyten, Lu Liu, Marie-­‐Francine Moens, Jie Tang, Juanzi Li Tsinghua Univeristy, Katholieke Universiteit Leuven, Northwestern University •  Diabetes are major causes of early death in most countries and over 68% of diabetes-­‐related mortality is caused by diabetes complica0ons. •  471 billion USD were spent on healthcare for 371 million diabetes pa0ents world-­‐wide in 2012, s0ll 4.8 million people died in 2012 due to diabetes. Input: a pa0ent’s lab test results Bilirubin example Output: diabetes complica0ons Routine urine analysis coronary heart disease diabe(c re(nopathy Proposed Model
Model the joint distribu0on of a given set of lab tests X over complica0on labels Y as Output Layer
Key challenge: feature sparseness Association vector
•  Averagely each clinical record only contains 1.26% of lab tests on average. The feature factor is defined as Latent Layer
•  65.5% types of lab tests are recorded in less than 0.0054% of clinical records. 0.1
0.2
0.4
Dimension reduction
0.2
0.1
Input Layer
0.5
/
/
0.3
/
0.6
/
/
The correla0on factor is defined as 0.4
The correla0on factor is defined as The model alleviates the sparseness issue by projec0ng feature space into a low-­‐dimensional latent space. Forecas0ng Performance
•  SVM and tradi0onal factor graph model suffer from the feature sparseness (-­‐59.9% compared with our method by recall) •  PCA improve the performance. However, it separates sparse coding and classifica0on into two processes. •  Our approach (SparseFGM) outperforms the baselines 20% by F1 •  Dataset: a collec0on of real medical records from a famous geriatric hospital: •  181,933 medical records, 35,525 pa0ents and 1,945 kinds of lab tests. •  On average each clinical record contains 24.43 lab tests (1.26% of all lab tests). •  We consider 3 complica0ons: •  hypertension (HTN), coronary heart disease (CHD), hyperlipidemia (HPL). r.
dep
DR
FL
.
ins
OP
bro
.
CV
D
L
HP
D
CH
HT
N
Associa0on Analysis
Vitamin C
KET
URO
BIL
WBC in the urine typically is found in urinary tract infec0ons which cause frequent voiding, which causes insomnia. Micro-­‐level: associa0on between 10 complica0ons and 9 parameters of rou0ne urine anlaysis discovered by the proposed mode.
Contact: [email protected]
0.59
OP
bro.
PRO
0.65
CVD
Nitrite
GLU
0.79
HPL
FL
Hyperlipidemia (HPL) can be diagnosed more precisely, while depression (depr.) is usually recognized from psychological inves0ga0on instead of physiological lab tests. 0.77
CHD
RBC
WBC
0.61
HTN
0.52
0.48
0.35
ins.
depr. 0.01
0
0.2
0.4
0.6
0.8
Recall
Macro-­‐level: study how each complica0on is diagnosable from lab test results. Code & Data: hGp://aminer.org/diabetes
Tsinghua University
Download