Data-driven Automatic Treatment Regimen Development and Recommendation Date: 2016/09/20 Author: Leilei Sun, Chuanren Liu, Chonghui Gou, Hui Xiong, Yatming Xie Source: ACM KDD’ 16 Advisor: Jia-ling Koh Speaker : Yi-hui Lee 1
Outline • Introduction • Approach • Experiment • Conclusion 2
Introduction • Motivation: The typical treatment regimens are usually used as prototypes when a clinical doctor designs the personalized treatment plan for a new patient. • Goal: Typical treatment regimens automatic identification and evaluation. 3
Introduction(cont.) • Treatment regimen recommendation framework: step 1 step 2 step 3 step 4 4
Introduction(cont.) • Data: EMRs: Electronic Medical Records • Challenges: Complex EMRs data: effective method to measure the similarity between treatment records Large volume of patients’ records: clustering algorithm->MRDPC 5
Outline • Introduction • Approach • Experiment • Conclusion 6
Approach • Treatment regimen recommendation framework: step 1 step 2 step 3 step 4 7 7
Approach(cont.) • Data: EMRs: Contain five categories of information of patients - Demographic information - Diagnostic information - Laboratory indicators - Doctor orders - Outcomes 8
Approach(cont.) • Demographic information: Age, gender, address, race and ethnicity, education, and other information of a patient. 9
Approach(cont.) • Diagnostic information(given by doctors): It consists of disease names and severity of the diseases. 10
Approach(cont.) • Laboratory indicators: Mainly used for judging the severity of disease or evaluating a patient’s final outcome, which can be implied by diagnostic information or the outcome, so which are not included in the following model. 11
Approach(cont.) • Treatments: Series of doctor orders • Doctor orders: Medical prescription Drug Name 12
Approach(cont.) • Treatments: Series of doctor orders • Doctor orders: Medical prescription Delivery Route IV: Intravenous injection IM: Intramuscular Per os, PO: Oral 13
Approach(cont.) • Treatments: Series of doctor orders • Doctor orders: Medical prescription Dosage each time 14
Approach(cont.) • Treatments: Series of doctor orders • Doctor orders: Medical prescription Frequency frequency per day 15
Approach(cont.) • Treatments: Series of doctor orders • Doctor orders: Medical prescription Repeating times during the q-th period 16
Approach(cont.) • Outcome: Evaluated and presented by doctors when a patient leaves hospital - Cured - Improved - Ineffective - Dead 17
Approach(cont.) • Similarity Measure for Treatment w = (0.4 0.2 0.2 0.2) • s ̄ q (i,j) is the similarity of OSP iq and OSP jq , which is the similarity of two treatments in q-th period. f iqg is the repetition times of g-th order in q-th period of treatment i {a ijqgh } is the allocation matrix for the computing of |OSP iq ∩ OSP jq | 18
Approach(cont.) • Similarity Measure for Treatment A S(A, D)=(1x[0+120/160])/2 B S(A, E)=(0x[1+24/160])/2 C DD= DD= S(B, D)=(0x[0+18/120])/2 DN DE DN DE O Dose × O Freq O Dose × O Freq … D 19 E
Approach(cont.) • Similarity Measure for Treatment • Properties: • 1) The value of s ̄ q (i,j) is from 0 to 1 • 2) Symmetry. For any i and j, s ̄ q (i,j) = s ̄ q (j,i); • 3)Self-similarity. s ̄ q (i,j)=1,if and only if OSP iq =OSP jq 20
Approach(cont.) • Clustering Treatments: Map Reduce enhanced Density Peaks based Clustering (MRDPC) • DPC: Density Peaks based Clustering - Derives from exemplar-based clustering algorithm - Discover clusters with complex shapes 21
Approach(cont.) • DPC Compute two indicators for each object χ (x)=1 if x>0 1) Local density ρ and χ (x)=0 otherwise meaning of ρ i is to count the number of objects in object i’s s c - neighborhood 2) Minimum distance (or maximum similarity) between the object and any other object with higher local density γ Objects with larger ρ and lower γ values are viewed as exemplars. 22
Approach(cont.) • MRDPC: The total N patients are first randomly divided into m parts Parts 1 Parts 1 … … p1 pN p1 Parts 2 p1 p2 p3 Parts 2 DPC … … p3 p9 p9 get k potential … p4 pN exemplars … … Parts m Parts m … … 23
Approach(cont.) • Extracting Typical Treatment Regimens: A typical treatment regimen includes the names of medicines used in a specified period, the dosages, the delivery routes, and lasting how many days. Extract a semantic description of each treatment cluster by its dense core. Dense core: Constructed by k-nearest neighbors of its exemplar. 24
Approach(cont.) • Extracting Typical Treatment Regimens: A treatment regimen Specified period λ (Drug, OSP ) = 1 if Drug is used in OSP (Drug ∈ OSP ), λ (Drug, OSP ) = 0 otherwise DA is a triple consists of delivery route, dosage and lasting days of a order 25
Approach(cont.) • Evaluating Typical Treatment Regimen: Patient cohort: The patients in a same leaf node is defined as a patient cohort - Decision tree model - Divide patients into different groups according to demographic information, diagnostic information and outcomes. 26
Approach(cont.) • Treatment regimen recommendation framework: step 1 step 2 step 3 step 4 27 27
Outline • Introduction • Approach • Experiment • Conclusion 28
Experiment • Experimental Data: EMRs data used in this paper are collected from Hos- pital Information Systems (HIS) of 14 Grade Three Class A (G3CA) hospitals Focus on the patients with cerebral infarction disease, which is one of the most common diseases in China today Extract the typical treatment regimens of cerebral infarction from doctor orders of 28, 659 patients. The total number of doctor orders is 1, 007, 057 29
Experiment(cont.) • Extracting Typical Treatment Regimens: • Select 138 medicines that are most relevant to cerebral infarction • 363,674 doctor orders containing the selected medicines, nearly 13 doctor orders per patient. 30
Experiment(cont.) • Extracting Typical Treatment Regimens: Four typical treatment regimens extracted from EMRs 31
Experiment(cont.) • Extracting Typical Treatment Regimens: An example of an extracted treatment regimen 32
Experiment(cont.) • Treatment Regimen Recommendation: Recommend treatment regimens for two patient cohorts 33
Experiment(cont.) • Treatment Regimen Recommendation: 34
Experiment(cont.) • Overall Treatment Evaluation: Our method can help improve effective rate and cure rate 35
Outline • Introduction • Approach • Experiment • Conclusion 36
Conclusion • Investigated how to identify the typical treatment regimens from large-scale treatment records and how to find the most effective treatment regimens for patients. Developed an efficient semantic clustering algorithm, based on a new method to measure the similarities between treatment records. Applied on large-scale treatment records, we were able to extract the treatment clusters as the typical treatment regimens with semantically meaningful descriptions. Designed a unified framework to evaluate the effectiveness of the identified treatment regimens • This work may be the first step towards the automatic development of treatment regimens and treatment recommendations. 37
Recommend
More recommend
