This document describes a study that used survival analysis techniques like the Kaplan-Meier estimator and Cox proportional hazards model to develop a predictive model for medication non-adherence using a large French healthcare database. The model analyzed factors like medical history, treatments, and healthcare utilization to estimate patient survival curves and risk scores for remaining on tamoxifen treatment for breast cancer. The results validated using these statistical techniques to predict medication abandonment risk and identified the most influential variables.
Independent Call Girls In Jaipur { 8445551418 } ✔ ANIKA MEHTA ✔ Get High Prof...
Predictive Patient Care: Survival model to prevent Medication Non-Adherence
1. Predictive Patient Care :
Survival Model to Prevent
Medication Non-Adherence
D. Kanoun, Clinique Pasteur
T. Janssoone, P. Rinder, P. Hornus, Sêmeia
2. • Cox proportional hazards model
• Survival analysis
• Iterative combination
• The studied Database: SNIIRAM
• Kaplan Meier estimator
2
Plan
Kaplan Meier
estimator
Cox proportional
hazards model
SNIIRAM
0 200 400 600 800 1000
0.00.20.40.60.81.0
Survival curves given the risk score
Number of days
Score < 0.8
0.8 ... score < 1.3
Score ... 1.3
3. 3
. The SNIIRAM Database: records of reimbursement data of the French health system.
. Focus on women’s breast cancer:
our cohort is composed of 50% of women, diagnosed with breast cancer, having purchased
Tamoxifen between 2013 and 2015.
. Raw data are processed to phase that allows the reconstruction of the patient’s care paths.
The Database
SNIIRAM
4. 0 200 400 600 800 1000
02468
Hazard function
Number of days
1/10000
. Kaplan-Meier Estimator: (or product-limit estimator ) probability of surviving in a given
length of time while considering time in many small intervals.
. Hazard function can be derivated from Kaplan-Meier estimator:
Roughly characterizes the « instantaneous probability » of a drug drop-out at time t
4
Survival Analysis
5. 5
. Cox regression:
estimates the effect of each characteristic to the study phenomena as an Hazard Ratio (HR)
HR = 1: no effect; HR < 1: reduction in hazard; HR > 1: increase in hazard
Proportional hazard model
Variable HR
CMU-C 1.47
ACS 1.52
Time since ALD status 1.08
Number of medical consultation 0.99
Psychiatric illness 1.19
Recent hospitalization for malignant
neoplasms of breast
0.68
Previous treatment: Tamoxifen 2.49
Previous treatment: radiotherapy 0.49
Previous treatment: chemotherapy 0.42
6. . Insight of Cox regression into Kaplan-Meier estimator:
Use of information about the patient to compute more accurate survival functions
For example:
1. Compute a risk score at the beginning of a phase
2. Survival function according to the previous treatment
1. Compute a risk score at the beginning of a phase
6
Combination
0 200 400 600 800 1000
0.00.20.40.60.81.0
Survival curves given the risk score
Number of days
Score < 0.8
0.8 ... score < 1.3
Score ... 1.3
0 200 400 600 800 1000
0.00.20.40.60.81.0
Survival curves given the previous treatment
Number of days
Chemotherapy
Radiotherapy
None
Tamoxifen
Other hormonotherapy
7. 7
• Validate the possibility to estimate the risk of abandonment during a phase of treatment
• Insights about the most significant variables and their weights.
• Limitations of the statistical models due to strong hypothesis
• Use in an application for patient support
• Test other approaches (deep learning, reinforcement learning, sequence mining, …)
• Provide a meta-analysis of these tools on different pathologies
Conclusion & Future work
8. Predictive Patient Care :
Survival Model to Prevent
Medication Non-Adherence
D. Kanoun, Clinique Pasteur
T. Janssoone, P. Rinder, P. Hornus, Sêmeia