Near-optimal Individualized Treatment Recommendations
Authors: Haomiao Meng, Ying-Qi Zhao, Haoda Fu, Xingye Qiao
JMLR 2020 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Simulation studies and a real data analysis for Type 2 diabetic patients with injectable antidiabetic treatments are conducted to show the usefulness of the proposed A-ITR framework. |
| Researcher Affiliation | Collaboration | Haomiao Meng EMAIL Department of Mathematical Sciences Binghamton University, State University of New York Binghamton, NY 13902, USA Ying-Qi Zhao EMAIL Public Health Sciences Division Fred Hutchinson Cancer Research Center Seattle, WA 98109, USA Haoda Fu EMAIL Eli Lilly and Company Indianapolis, IN 46285, USA Xingye Qiao EMAIL Department of Mathematical Sciences Binghamton University, State University of New York Binghamton, NY 13902, USA |
| Pseudocode | Yes | 4.1 Algorithm In this section, we introduce the optimization procedure to estimate f defined in (6). Instead of the constrained problem (7), we solve the regularized problem: min f F 1 n Σi=1 yi p(ai|xi)ℓ( Wai, f(xi) ) + λJ(f), (10) where λ is a tuning parameter. It is a weighted classification problem with weight wi = yi/p(ai|xi). ... We denote the coefficient matrix as Bp (k 1) = [β1, . . . , βk 1]. Then we create another copy of the coefficients Gp (k 1) = [γ1, . . . , γk 1], and let Zp (k 1) = [z1, . . . , zk 1]. Recall that wi = yi/p(ai|xi), then we minimize the augmented Lagrangian Lρ(B, G, Z) = Σi=1 wiℓ1( Wai, BT xi ) + Σi=1 wiℓ2( Wai, GT xi ) + nλ Σj=1 βT j βj + Σj=1 zT j (βj γj) + ρ/2 Σj=1 (βj γj)T (βj γj), where ρ > 0 controls the step size. At step t, for each j = 1, . . . , k 1 we can update Bt, Gt and Zt as βt j = argmin βj Lρ([βt 1, . . . , βj, . . . , βt 1 k 1], Gt 1, Zt 1), γt j = argmin γj Lρ(Bt, [γt 1, . . . , γj, . . . , γt 1 k 1], Zt 1), zt j = zt 1 j + ρ(βt j γt j) until matrix B converges. |
| Open Source Code | Yes | An R package aitr has been developed, found at https://github.com/menghaomiao/aitr. |
| Open Datasets | No | The paper mentions a "Type 2 diabetes mellitus (T2DM) observational study" and states "The data set contains 1139 patients." However, it does not provide any link, DOI, or formal citation to a publicly accessible repository for this specific dataset. |
| Dataset Splits | Yes | In each case, we let the training sample size to be n = 500, 1000, 2000, and use a test set with sample size 1000 to evaluate the performance. ...we split the data into a training set (70%) and a test set (30%). |
| Hardware Specification | No | The paper does not specify any particular hardware (e.g., GPU, CPU models, memory) used for running the experiments or simulations. |
| Software Dependencies | No | The paper mentions that "An R package aitr has been developed" and uses "logistic regression" and "ordinary least square regression". However, it does not provide specific version numbers for R, the aitr package, or any other software libraries or tools used for the analysis. |
| Experiment Setup | Yes | The first step is to tune λ. For each λ, the estimated solution is ˆf. Then we define the corresponding single-treatment ITR as d ˆ f = argmaxj Wj, ˆf and calculate its empirical average of the expected outcome (1)... We choose the λ that yields the smallest empirical risk for the resulting ITR... For the one-step method, we need to continue to tune δ. Example 1: ...The tuning parameter λ is chosen from 5^-6 to 5^2. Example 2: ...The tuning parameter λ is chosen from 5^-9 to 5^-1. Example 3: ...The tuning parameter λ is chosen from 5^-7 to 1. |