AI helps you reading Science

AI generates interpretation videos

AI extracts and analyses the key points of the paper to generate videos automatically

Go Generating

AI Traceability

AI parses the academic lineage of this thesis

Master Reading Tree
Generate MRT

AI Insight

AI extracts a summary of this paper

We present two case studies where high-performance generalized additive models with pairwise interactions are applied to real healthcare problems yielding intelligible models with state-of-the-art accuracy

Intelligible Models for HealthCare: Predicting Pneumonia Risk and Hospital 30-day Readmission

ACM Knowledge Discovery and Data Mining, pp.1721-1730, (2015)

Cited by: 864|Views410


In machine learning often a tradeoff must be made between accuracy and intelligibility. More accurate models such as boosted trees, random forests, and neural nets usually are not intelligible, but more intelligible models such as logistic regression, naive-Bayes, and single decision trees often have significantly worse accuracy. This tra...More



  • Sections 3 and 4 present the case studies of training intelligible GA2M model on the pneumonia and the 30-day readmission data, respectively.
  • # outpatient visits ever predictive terms in the 30-day readmission model measure the number of visits patients have made in the last 12 month, 6 months, and 3 months to the ER, as an outpatient, and as an inpatient.
  • In this paper we describe the application of a learning method based on high-performance generalized additive models [5, 6] to the pneumonia problem described above, and to a modern, much larger problem predicting 30-day hospital readmission
  • The term risk scores are added to a baseline risk, and the sum converted to a probability
  • Patients with aggregate risk scores above -2.2 have higher than average risk, and patients with total risk scores below -2.2 have lower than average risk scores
  • We present two case studies on real medical data where GA2Ms achieve state-of-the-art accuracy while remaining intelligible
  • We believe GA2Ms represent a significant step forward in the tradeoff between model accuracy and intelligibility that should make it easier to deploy high-accuracy learned models in applications such as healthcare where model verification and debuggability are as important as accuracy
  • The main reason this patient is predicted to be likely to return is because they have been in the hospital often in the last year, but the first few terms in the model give them a hint of the medical conditions that put them at elevated risk.
  • The most important 6 terms are preparations that the patient received during their last visit: prednisone is a corticosteroid used as an imummosuppressant, etoposide in an anticancer drug, mesna is a cancer chemotherapy drug, doxorubicin is a treatment for blood and skin cancers, dexamethosone is another immunosuppressant steroid, and ondansetron is a drug used to treat nausea from chemotherapy.
  • This 6 terms that increase this patient’s readmission risk the most are: 1) the patient has endrometrial carcinoma; 2) a benign abdominal tumor; 3) a relaxant typically prescribed to calm patients or reduce spasms; 4) a fairly typical hematocrit test result; 5) a precancerous non-invasive lesion in the breast; and 6) a small number of outpatient visits suggesting they have been receiving treatment as an outpatient without needing to be hospitalized.
  • Because the prediction task is hospital readmission, not probability of death, age represents a weaker, more generic characterization of patient health and their likeliness to need additional hospitalization within 30 days.
  • One possible explanation for this is that in a dataset from the 90’s, many patients would have retired at around age 65, and that this may yield differences in activity levels, health insurance, and willingness to get access healthcare early enough to improve outcomes — pneumonia responds well to treatment with antibiotics, but can be life threatening if not treated.
  • Either medical treatments are effective for patients older than 85, or other medical conditions are more likely to be responsible for death at this age than pneumonia, or risk does increase above 85 and the model has failed to learn it.
  • More complex 30-day hospital readmission task the GA2M model achieves excellent accuracy while yielding a manageable, surprisingly intelligible model despite incorporating over 4000 terms.
  • The authors believe GA2Ms represent a significant step forward in the tradeoff between model accuracy and intelligibility that should make it easier to deploy high-accuracy learned models in applications such as healthcare where model verification and debuggability are as important as accuracy
  • Table1: Pneumonia attributes, grouped by type. Continuous features that will be shaped by GAM/GA2M models are marked with a “C”
  • Table2: AUC for different learning methods on the pneumonia and 30-day readmission tasks
  • Table3: Risk scores (log odds) and the corresponding probabilities
Download tables as Excel
  • In the mid 90’s, a large multi-institutional project was funded by Cost-Effective HealthCare (CEHC) to evaluate the application of machine learning to important problems in healthcare such as predicting pneumonia risk
  • R. Ambrosino, B. Buchanan, G. Cooper, and M. Fine. The use of misclassification costs to learn rule-based decision support models for cost-effective hospital admission strategies. In Proceedings of the Annual Symp. on Comp. Application in Medical Care, 1995.
    Google ScholarLocate open access versionFindings
  • G. Cooper, V. Abraham, C. Aliferis, J. Aronis, B. Buchanan, R. Caruana, M. Fine, J. Janosky, G. Livingston, T. Mitchell, S. Montik, and P. Spirtes. Predicting dire outcomes of patients with community acquired pneumonia. Journal of Biomedical Informatics, 38(5):347–366, 2005.
    Google ScholarLocate open access versionFindings
  • G. Cooper, C. Aliferis, R. Ambrosino, J. Aronis, B. Buchanan, R. Caruana, M. Fine, C. Glymour, G. Gordon, B. Hanusa, J. Janosky, C. Meek, T. Mitchell, T. Richardson, and P. Spirtes. An evaluation of machine-learning methods for predicting pneumonia mortality. Artificial Intelligence in Medicine, 9(2):107–138, 1997.
    Google ScholarLocate open access versionFindings
  • T. Hastie and R. Tibshirani. Generalized additive models. Chapman & Hall/CRC, 1990.
    Google ScholarFindings
  • Y. Lou, R. Caruana, and J. Gehrke. Intelligible models for classification and regression. In KDD, 2012.
    Google ScholarLocate open access versionFindings
  • Y. Lou, R. Caruana, J. Gehrke, and G. Hooker. Accurate intelligible models with pairwise interactions. In KDD, 2013.
    Google ScholarFindings
  • S. Wood. Generalized additive models: an introduction with R. CRC Press, 2006.
    Google ScholarFindings
Your rating :