Healthcare expenditure is set to rise over the coming years. Cost will undoubtedly influence patients’ decision-making when it comes to diagnosis and treatment.
For healthcare providers, providing up-front cost estimates improves patient experience, making patients more willing to return (if required) in the future. For patients, having accurate pre-admission estimates allow for informed decisions and adequate preparation, reducing payment challenges after treatment. Ultimately, this case is a first step towards (i) standardization of healthcare cost estimation and (ii) price transparency to build trust between healthcare providers, payers, and patients.
1. HOW UCARE.AI PREDICTS
HEALTHCARE COSTS
To improve patient experience and reduce payment challenges
DATAx Singapore, 5th March 2019
2. Koh D. (2018, December 19). Parkway Pantai hospitals launch AI-powered predictive hospital
bill estimation system, Healthcare IT News. Retrieved from https://www.healthcareitnews.com
4. Current health expenditure per capita, PPP (current international $), World
Bank, Retrieved from https://data.worldbank.org on 25 Feb 2019
Healthcare costs have risen, and will rise more in the future
5. Baker, J.A. (2018, April 07). Singapore ranks high in report on
medical inflation in Asia, Channel News Asia. Retrieved from
https://www.channelnewsasia.com
6. (2018, November 18). Medical inflation in Singapore to hit 10% in 2019,
Singapore Business Review. Retrieved from https://sbr.com.sg
7. 2019 Global Medical Trend Rates Report, Aon.
Retrieved from https://www.aon.com on 12 Feb 2018
8. Survey finds healthcare providers benefit from up-front cost
estimates, yet many patients find it difficult to secure such
information, TransUnion. Retrieved from
https://www.newsroom.transunion on 25 Feb 2018
For Providers:
Up-front cost estimates improve patients experience and retention
10. INTENT
Provide potential patients with (more) accurate estimates
Preference for overestimation, instead of underestimation
Minimal disruption (i.e., plug-n-play) <- Constraint
15. DATA VALIDATION & INGESTION
Encrypt
Decrypt
Data Validation
- Schema, format checks
- Duplicate, null value
checks
- Volume checks
Provider
environment
uCare
environment
Extract
Load
Transform
16. DATA PREPARATION
* International Classification of Diseases
Categoricals
- Converting ICD9 to ICD10*
- Standardization (e.g., gender, room)
Numerics
- Filling missing values
- Handling outliers
Etc. (e.g., text, images, EHRs)
Merging across
multiple datasets
Data Augmentation
External data (e.g.,
doctor data from SMC)
Internal features (e.g.,
disease embeddings)
17. VALIDATION
* (K-fold) Cross Validation
Training fold
Validation fold
- Shuffle data (randomly)
- Split into k-folds and cross
validate
Training fold
Validation fold
- Sort data by time (e.g.,
admission date)
- Split into time periods (e.g.,
yearly, half-yearly)
- Train only on data from
period(s) before
Validation
- Random split/CV*
- Time-based split/CV
Train set
Validation set
18. VALIDATION (is so important there are two slides!)
- 19% of patients had >1
admission (38% of cases)
- In validation, >100 cases of
ICD10/TOSP not seen before
- Medical prices increase over time
TIME
Cannot use future to predict past—validation results differ greatly
23. MISCELLANEOUS TIPS
Take time to learn from domain experts and users
More data and/or features != Better model
Proper engineering practices make (everyone’s) life easier
24. OVERALL OUTCOMES
Accuracy and usability improvements
- Reduced MAE (55%), RMSE (60%), and underestimation %
- Improved query latency (< 1 sec)
- Customization at institution level
Additional features
- Integration of business logic layer for strategic purposes
25. OVERALL OUTCOMES
Minimal disruption to front-line users
- API-based, existing front-end used—no training required
Reduced cost, increased frequency of updates
- System built for daily updates
Improved user & customer experience
- Based on anecdotal evidence
26. KEY TAKEAWAYS
Building useful data products is a team effort
Be humble of the data—don’t assume anything
Machine learning is < 20% of effort—the methodology and
engineering process is more important