This repository provides utilities and scripts to run limited automatic tabular ML pipelines for generic MEDS datasets.
This repository consists of two key pieces:
To use MEDS-Tab, install the dependencies following commands below. Note that this version of MEDS-Tab is compatible with MEDS v0.3
Pip Install
pip install meds-tab
Local Install
# clone the git repo
pip install .
For an end to end example, including re-sharding the input via MEDS-Transforms, see this example script
See /tests/test_integration.py
for a local example of the end-to-end pipeline (minus re-sharding) being run on synthetic data. This script is a functional test that is also run with pytest
to verify the correctness of the algorithm.
MEDS-Tab is a comprehensive framework designed to streamline the handling, modeling, and analysis of complex medical time-series data. By leveraging automated processes, MEDS-Tab significantly reduces the computation required to generate high-quality baseline models for diverse supervised learning tasks.
MEDS-Tab leverages the recently developed, minimal, easy-to-use Medical Event Data Standard (MEDS) schema to standardize structured EHR data to a consistent schema from which baselines can be reliably produced across arbitrary tasks and settings. In order to use MEDS-Tab, you will first need to transform your raw EHR data to a MEDS format, which can be done using the following libraries:
Run the MEDS-Tab Command-Line Interface tool (MEDS-Tab-cli
) to extract cohorts based on your task - check out the Usage Guide!
Painless Reproducibility: Use MEDS-Tab to obtain comparable, reproducible, and well-tuned XGBoost results tailored to your dataset-specific feature space!
By following these steps, you can seamlessly transform your dataset, define necessary criteria, and leverage powerful machine learning tools within the MEDS-Tab ecosystem. This approach not only simplifies the process but also ensures high-quality, reproducible results for your machine learning tasks for health projects. It can reliably take no more than a week of full-time human effort to perform Steps I-V on new datasets in reasonable raw formulations!
First, if your data is not already sharded to the degree you want and in a manner that subdivides your
splits with the format "$SPLIT_NAME/\d+.parquet"
, where $SPLIT_NAME
does not contain slashes, you will
need to re-shard your data. This can be done via the
MEDS-Transforms library, which is not included in this
repository. Having data sharded by split is a necessary step to ensure that the data is efficiently
processed in parallel. You can easily re-shard your input MEDS cohort in the environment into which this
package is installed with the following command:
# Re-shard pipeline
# $MIMICIV_MEDS_DIR is the directory containing the input, MEDS v0.3 formatted MIMIC-IV data
# $MEDS_TAB_COHORT_DIR is the directory where the re-sharded MEDS dataset will be stored, and where your model
# will store cached files during processing by default.
# $N_PATIENTS_PER_SHARD is the number of patients per shard you want to use.
MEDS_transform-reshard_to_split \
input_dir="$MIMICIV_MEDS_DIR" \
cohort_dir="$MEDS_TAB_COHORT_DIR" \
'stages=["reshard_to_split"]' \
stage="reshard_to_split" \
stage_configs.reshard_to_split.n_patients_per_shard=$N_PATIENTS_PER_SHARD
meds-tab-describe
: This command processes MEDS data shards to compute the frequencies of different code types. It differentiates codes into the following categories:
This script further caches feature names and frequencies in a dataset stored in a code_metadata.parquet
file within the MEDS_cohort_dir
argument specified as a hydra-style command line argument.
meds-tab-tabularize-static
: Filters and processes the dataset based on the frequency of codes, generating a tabular vector for each patient at each timestamp in the shards. Each row corresponds to a unique patient_id
and timestamp
combination, thus rows are duplicated across multiple timestamps for the same patient.
Example: Tabularizing static data with the minimum code frequency of 10, window sizes of [1d, 30d, 365d, full]
, and value aggregation methods of [static/present, static/first, code/count, value/count, value/sum, value/sum_sqd, value/min, value/max]
meds-tab-tabularize-static MEDS_cohort_dir="path_to_data" \
tabularization.min_code_inclusion_frequency=10 \
tabularization.window_sizes=[1d,30d,365d,full] \
do_overwrite=False \
tabularization.aggs=[static/present,static/first,code/count,value/count,value/sum,value/sum_sqd,value/min,value/max]"
/src/MEDS_tabular_automl/utils.py
meds-tab-tabularize-time-series
: Iterates through combinations of a shard, window_size
, and aggregation
to generate feature vectors that aggregate patient data for each unique patient_id
x timestamp
. This stage (and the previous stage) uses sparse matrix formats to efficiently handle the computational and storage demands of rolling window calculations on large datasets. We support parallelization through Hydra's --multirun
flag and the joblib
launcher.
Example: Aggregate time-series data on features across different window_sizes
meds-tab-tabularize-time-series --multirun \
worker="range(0,$N_PARALLEL_WORKERS)" \
hydra/launcher=joblib \
MEDS_cohort_dir="path_to_data" \
tabularization.min_code_inclusion_frequency=10 \
do_overwrite=False \
tabularization.window_sizes=[1d,30d,365d,full] \
tabularization.aggs=[static/present,static/first,code/count,value/count,value/sum,value/sum_sqd,value/min,value/max]
meds-tab-cache-task
: Aligns task-specific labels with the nearest prior event in the tabularized data. It requires a labeled dataset directory with three columns (patient_id
, timestamp
, label
) structured similarly to the MEDS_cohort_dir
.
Example: Align tabularized data for a specific task $TASK
and labels that has pulled from ACES
meds-tab-cache-task MEDS_cohort_dir="path_to_data" \
task_name=$TASK \
tabularization.min_code_inclusion_frequency=10 \
do_overwrite=False \
tabularization.window_sizes=[1d,30d,365d,full] \
tabularization.aggs=[static/present,static/first,code/count,value/count,value/sum,value/sum_sqd,value/min,value/max]
meds-tab-xgboost
: Trains an XGBoost model using user-specified parameters. Permutations of window_sizes
and aggs
can be generated using generate-subsets
command (See the section below for descriptions).
meds-tab-xgboost --multirun \
MEDS_cohort_dir="path_to_data" \
task_name=$TASK \
output_dir="output_directory" \
tabularization.min_code_inclusion_frequency=10 \
tabularization.window_sizes=$(generate-subsets [1d,30d,365d,full]) \
do_overwrite=False \
tabularization.aggs=$(generate-subsets [static/present,static/first,code/count,value/count,value/sum,value/sum_sqd,value/min,value/max])
generate-subsets
: Generates and prints a sorted list of all non-empty subsets from a comma-separated input. This is provided for the convenience of sweeping over all possible combinations of window sizes and aggregations.
For example, you can directly call generate-subsets
in the command line:
generate-subsets [2,3,4] \
[2], [2, 3], [2, 3, 4], [2, 4], [3], [3, 4], [4]
This could be used in the command line in concert with other calls. For example, the following call:
meds-tab-xgboost --multirun tabularization.window_sizes=$(generate-subsets [1d,2d,7d,full])
would resolve to:
meds-tab-xgboost --multirun tabularization.window_sizes=[1d],[1d,2d],[1d,2d,7d],[1d,2d,7d,full],[1d,2d,full],[1d,7d],[1d,7d,full],[1d,full],[2d],[2d,7d],[2d,7d,full],[2d,full],[7d],[7d,full],[full]
which can then be correctly interpreted by Hydra's multirun logic.
MEDS-Tab has several key limitations which we plan to address in future changes. These include, and are tracked by, the following GitHub issues.
This is a common misconception. Tabular data refers to data that can be organized in a consistent, logical set of rows/columns such that the entirety of a "sample" or "instance" for modeling or analysis is contained in a single row, and the set of columns possibly observed (there can be missingness) is consistent across all rows. Structured EHR data does not satisfy this definition, as we will have different numbers of observations of medical codes and values at different timestamps for different patients, so it cannot simultanesouly satisfy the (1) "single row single instance", (2) "consistent set of columns", and (3) "logical" requirements. Thus, in this pipeline, when we say we will produce a "tabular" view of MEDS data, we mean a dataset that can realize these constraints, which will explicitly involve summarizing the patient data over various historical or future windows in time to produce a single row per patient with a consistent, logical set of columns (though there may still be missingness).
In this section, we describe the MEDS-Tab architecture, specifically some of the pipeline choices we made to reduce memory usage and increase speed during the tabularization process and XGBoost tuning process.
We break our method into 4 discrete parts:
This initial stage processes a pre-shareded dataset. We expect a structure as follows where each shard contains a subset of the patients:
/PATH/TO/MEDS/DATA
│
└─── <SPLIT A>
│ │ <SHARD 0>.parquet
│ │ <SHARD 1>.parquet
│ │ ...
│
└─── <SPLIT B>
│ │ <SHARD 0>.parquet
│ │ <SHARD 1>.parquet
| │ ...
|
...
We then compute and store feature frequencies, crucial for determining which features are relevant for further analysis.
Detailed Workflow:
The tabularization stage of our pipeline, exposed via the cli commands:
meds-tab-tabularize-static
for tabularizing static datameds-tab-tabularize-time-series
for tabularizing the time series dataStatic data is relatively small in the medical datasets, so we use a dense pivot operation, convert it to a sparse matrix, and then duplicate rows such that the static data will match up with the time series data rows generated in the next step. Static data is currently processed serially.
The script for tabularizing time series data primarily transforms a raw, unstructured dataset into a structured, feature-rich dataset by utilizing a series of sophisticated data processing steps. This transformation (as depicted in the figure below) involves converting raw time series from a Polars dataframe into a sparse matrix format, aggregating events that occur at the same date for the same patient, and then applying rolling window aggregations to extract temporal features.
Data Loading and Categorization:
Sparse Matrix Conversion:
Rolling Window Aggregation:
Output Storage:
.npz
file on disk./PATH/TO/MEDS/TABULAR_DATA
│
└─── <SPLIT A>
├─── <SHARD 0>
│ ├───code
│ │ └───count.npz
│ └───value
│ └───sum.npz
...
Now that we have generated tabular features for all the events in our dataset, we can cache subsets relevant for each task we wish to train a supervised model on. This step is critical for efficiently training machine learning models on task-specific data without having to load the entire dataset.
Detailed Workflow:
The file structure for the cached data mirrors that of the tabular data, also consisting of .npz
files, where users must specify the directory that stores labels. Labels follow the same shard file structure as the input meds data from step (1), and the label parquets need patient_id
, timestamp
, and label
columns.
The final stage uses the processed and cached data to train an XGBoost model. This stage is optimized to handle the sparse data structures produced in earlier stages efficiently.
Detailed Workflow:
Evaluating the computational overhead of tabularization methods is essential for assessing their efficiency and suitability for large-scale medical data processing. This section presents a comparative analysis of the computational overhead of MEDS-Tab with other systems like Catabra and TSFresh. It outlines the performance of each system in terms of wall time, memory usage, and output size, highlighting the computational efficiency and scalability of MEDS-Tab.
The systems compared in this study represent different approaches to data tabularization, with the main difference being MEDS-Tab usage of sparse tabularization. Specifically, for comparison we used:
The benchmarking tests were conducted using the following hardware and software settings:
Tabularization of time-series data, as depicted above, is commonly used in several past works. The only two libraries to our knowledge that provide a full tabularization pipeline are tsfresh
and catabra
. catabra
also offers a slower but more memory-efficient version of their method which we denote catabra-mem
. Other libraries either provide only rolling window functionalities (featuretools
) or just pivoting operations (Temporai
/Clairvoyance
, sktime
, AutoTS
). We provide a significantly faster and more memory-efficient method. Our findings show that on the MIMIC-IV and eICU medical datasets, we significantly outperform both above-mentioned methods that provide similar functionalities with MEDS-Tab. While catabra
and tsfresh
could not even run within a budget of 10 minutes on as low as 10 patients' data for eICU, our method scales to process hundreds of patients with low memory usage under the same time budget. We present the results below.
The tables below detail computational resource utilization across two datasets and various patient scales, emphasizing the better performance of MEDS-Tab in all of the scenarios. The tables are organized by dataset and number of patients. For the analysis, the full window sizes and the aggregation method code_count were used. Additionally, we use a budget of 10 minutes for running our tests given that for such a small number of patients (10, 100, and 500 patients) data should be processed quickly. Note that catabra-mem
is omitted from the tables as it was never completed within the 10-minute budget.
The only method that was able to tabularize eICU data was MEDS-Tab. We ran our method with both 100 and 500 patients, resulting in an increment of three times in the number of codes. MEDS-Tab gave efficient results in terms of both time and memory usage.
a) 100 Patients
Table 1: 6,374 Codes, 2,065,608 Rows, Output Shape [132,461, 6,374]
Wall Time | Avg Memory | Peak Memory | Output Size | Method |
---|---|---|---|---|
0m39s | 5,271 MB | 14,791 MB | 362 MB | meds_tab |
b) 500 Patients
Table 2: 18,314 Codes, 8,737,355 Rows, Output Shape [565,014, 18,314]
Wall Time | Avg Memory | Peak Memory | Output Size | Method |
---|---|---|---|---|
3m4s | 8,335 MB | 15,102 MB | 1,326 MB | meds_tab |
MEDS-Tab, tsfresh
, and catabra
were tested across three different patient scales on MIMIC-IV.
a) 10 Patients
This table illustrates the efficiency of MEDS-Tab in processing a small subset of patients with extremely low computational cost and high data throughput, outperforming tsfresh
and catabra
in terms of both time and memory efficiency.
Table 3: 1,504 Codes, 23,346 Rows, Output Shape [2,127, 1,504]
Wall Time | Avg Memory | Peak Memory | Output Size | Method |
---|---|---|---|---|
0m2s | 423 MB | 943 MB | 7 MB | meds_tab |
1m41s | 84,159 MB | 265,877 MB | 1 MB | tsfresh |
0m15s | 2,537 MB | 4,781 MB | 1 MB | catabra |
b) 100 Patients
The performance gap was further highlighted with an increased number of patients and codes. For a moderate patient count, MEDS-Tab demonstrated superior performance with significantly lower wall times and memory usage compared to tsfresh
and catabra
.
Table 4: 4,154 Codes, 150,789 Rows, Output Shape [15,664, 4,154]
Wall Time | Avg Memory | Peak Memory | Output Size | Method |
---|---|---|---|---|
0m5s | 718 MB | 1,167 MB | 45 MB | meds_tab |
5m9s | 217,477 MB | 659,735 MB | 4 MB | tsfresh |
3m17s | 14,319 MB | 28,342 MB | 4 MB | catabra |
c) 500 Patients
Scaling further to 500 patients, MEDS-Tab maintained consistent performance, reinforcing its capability to manage large datasets efficiently. Because of the set time limit of 10 minutes, we could not get results for catabra
and tsfresh
. In comparison, MEDS-Tab processed the data in about 15 seconds, making it at least 40 times faster for the given patient scale.
Table 5: 48,115 Codes, 795,368 Rows, Output Shape [75,595, 8,115]
Wall Time | Avg Memory | Peak Memory | Output Size | Method |
---|---|---|---|---|
0m16s | 1,410 MB | 3,539 MB | 442 MB | meds_tab |
Evaluating our tabularization approach for baseline models involved training XGBoost across a spectrum of binary clinical prediction tasks, using data from the MIMIC-IV database. These tasks encompassed diverse outcomes such as mortality predictions over different intervals, readmission predictions, and lengths of stay (LOS) in both ICU and hospital settings.
Each task is characterized by its specific label and prediction time. For instance, predicting "30-day readmission" involves assessing whether a patient returns to the hospital within 30 days, with predictions made at the time of discharge. This allows input features to be derived from the entire duration of the patient's admission. In contrast, tasks like "In ICU Mortality" focus on predicting the occurrence of death using only data from the first 24 or 48 hours of ICU admission. Specifically, we use the terminology "Index Timestamp" to mean the timestamp such that no event included as input will occur later than this point.
We optimize predictive accuracy and model performance by using varied window sizes and aggregations of patient data. This approach allows us to effectively capture and leverage the temporal dynamics and clinical nuances inherent in each prediction task.
A single XGBoost run was completed to profile time and memory usage. This was done for each $TASK
using the following command:
meds-tab-xgboost
MEDS_cohort_dir="path_to_data" \
task_name=$TASK \
output_dir="output_directory" \
do_overwrite=False \
This uses the default minimum code inclusion frequency, window sizes, and aggregations from the launch_xgboost.yaml
:
allowed_codes: # allows all codes that meet min code inclusion frequency
min_code_inclusion_frequency: 10
window_sizes:
- 1d
- 7d
- 30d
- 365d
- full
aggs:
- static/present
- static/first
- code/count
- value/count
- value/sum
- value/sum_sqd
- value/min
- value/max
Since this includes every window size and aggregation, it is the most expensive to run. The runtimes and memory usage are reported below.
Task | Index Timestamp | Real Time | User Time | Sys Time | Avg Memory (MiB) | Peak Memory (MiB) |
---|---|---|---|---|---|---|
Post-discharge 30 day Mortality | Discharge | 2m59s | 3m38s | 0m38s | 9,037 | 11,955 |
Post-discharge 1 year Mortality | Discharge | 5m16s | 6m10s | 0m59s | 10,804 | 12,330 |
30 day Readmission | Discharge | 2m30s | 3m3s | 0m39s | 13,199 | 18,677 |
In ICU Mortality | Admission + 24 hr | 0m38s | 1m3s | 0m13s | 1,712 | 2,986 |
In ICU Mortality | Admission + 48 hr | 0m34s | 1m1s | 0m13s | 1,613 | 2,770 |
In Hospital Mortality | Admission + 24 hr | 2m8s | 2m41s | 0m32s | 9,072 | 12,056 |
In Hospital Mortality | Admission + 48 hr | 1m54s | 2m25s | 0m29s | 8,858 | 12,371 |
LOS in ICU > 3 days | Admission + 24 hr | 2m3s | 2m37s | 0m28s | 4,650 | 5,715 |
LOS in ICU > 3 days | Admission + 48 hr | 1m44s | 2m18s | 0m24s | 4,453 | 5,577 |
LOS in Hospital > 3 days | Admission + 24 hr | 6m5s | 7m5s | 1m4s | 11,012 | 12,223 |
LOS in Hospital > 3 days | Admission + 48 hr | 6m10s | 7m12s | 1m4s | 10,703 | 11,830 |
To better understand the runtimes, we also report the task specific cohort size.
Task | Index Timestamp | Number of Patients | Number of Events |
---|---|---|---|
Post-discharge 30 day Mortality | Discharge | 149,014 | 356,398 |
Post-discharge 1 year Mortality | Discharge | 149,014 | 356,398 |
30 day Readmission | Discharge | 17,418 | 377,785 |
In ICU Mortality | Admission + 24 hr | 7,839 | 22,811 |
In ICU Mortality | Admission + 48 hr | 6,750 | 20,802 |
In Hospital Mortality | Admission + 24 hr | 51,340 | 338,614 |
In Hospital Mortality | Admission + 48 hr | 47,231 | 348,289 |
LOS in ICU > 3 days | Admission + 24 hr | 42,809 | 61,342 |
LOS in ICU > 3 days | Admission + 48 hr | 42,805 | 61,327 |
LOS in Hospital > 3 days | Admission + 24 hr | 152,126 | 360,208 |
LOS in Hospital > 3 days | Admission + 48 hr | 152,120 | 359,020 |
The XGBoost sweep was run using the following command for each $TASK
:
meds-tab-xgboost --multirun \
MEDS_cohort_dir="path_to_data" \
task_name=$TASK \
output_dir="output_directory" \
tabularization.window_sizes=$(generate-subsets [1d,30d,365d,full]) \
do_overwrite=False \
tabularization.aggs=$(generate-subsets [static/present,code/count,value/count,value/sum,value/sum_sqd,value/min,value/max])
The model parameters were set to:
model:
booster: gbtree
device: cpu
nthread: 1
tree_method: hist
objective: binary:logistic
The hydra sweeper swept over the parameters:
params:
+model_params.model.eta: tag(log, interval(0.001, 1))
+model_params.model.lambda: tag(log, interval(0.001, 1))
+model_params.model.alpha: tag(log, interval(0.001, 1))
+model_params.model.subsample: interval(0.5, 1)
+model_params.model.min_child_weight: interval(1e-2, 100)
+model_params.model.max_depth: range(2, 16)
model_params.num_boost_round: range(100, 1000)
model_params.early_stopping_rounds: range(1, 10)
tabularization.min_code_inclusion_frequency: tag(log, range(10, 1000000))
Note that the XGBoost command shown includes tabularization.window_sizes
and tabularization.aggs
in the parameters to sweep over.
For a complete example on MIMIC-IV and for all of our config files, see the MIMIC-IV companion repository.
Task | Index Timestamp | AUC | Minimum Code Inclusion Frequency | Number of Included Codes* | Window Sizes | Aggregations |
---|---|---|---|---|---|---|
Post-discharge 30 day Mortality | Discharge | 0.935 | 1,371 | 5,712 | [7d,full] | [code/count,value/count,value/min,value/max] |
Post-discharge 1 year Mortality | Discharge | 0.898 | 289 | 10,048 | [2h,12h,1d,30d,full] | [static/present,code/count,value/sum_sqd,value/min] |
30 day Readmission | Discharge | 0.708 | 303 | 9,903 | [30d,365d,full] | [code/count,value/count,value/sum,value/sum_sqd,value/max] |
In ICU Mortality | Admission + 24 hr | 0.661 | 7,059 | 3,037 | [12h,full] | [static/present,code/count,value/sum,value/min,value/max] |
In ICU Mortality | Admission + 48 hr | 0.673 | 71 | 16,112 | [1d,7d,full] | [static/present,code/count,value/sum,value/min,value/max] |
In Hospital Mortality | Admission + 24 hr | 0.812 | 43 | 18,989 | [1d,full] | [static/present,code/count,value/sum,value/min,value/max] |
In Hospital Mortality | Admission + 48 hr | 0.810 | 678 | 7,433 | [1d,full] | [static/present,code/count,value/count] |
LOS in ICU > 3 days | Admission + 24 hr | 0.946 | 30,443 | 1,624 | [2h,7d,30d] | [static/present,code/count,value/count,value/sum,value/sum_sqd,value/max] |
LOS in ICU > 3 days | Admission + 48 hr | 0.967 | 2,864 | 4,332 | [2h,7d,30d] | [code/count,value/sum_sqd,value/max] |
LOS in Hospital > 3 days | Admission + 24 hr | 0.943 | 94,633 | 912 | [12h,1d,7d] | [code/count,value/count,value/sum_sqd] |
LOS in Hospital > 3 days | Admission + 48 hr | 0.945 | 30,880 | 1,619 | [1d,7d,30d] | [code/count,value/sum,value/min,value/max] |
Additionally, the model parameters from the highest-performing run are reported below.
Task | Index Timestamp | Eta | Lambda | Alpha | Subsample | Minimum Child Weight | Number of Boosting Rounds | Early Stopping Rounds | Max Tree Depth |
---|---|---|---|---|---|---|---|---|---|
Post-discharge 30 day Mortality | Discharge | 0.006 | 0.032 | 0.374 | 0.572 | 53 | 703 | 9 | 16 |
Post-discharge 1 year Mortality | Discharge | 0.009 | 0.086 | 0.343 | 0.899 | 76 | 858 | 9 | 11 |
30 day Readmission | Discharge | 0.006 | 0.359 | 0.374 | 0.673 | 53 | 712 | 9 | 16 |
In ICU Mortality | Admission + 24 hr | 0.038 | 0.062 | 0.231 | 0.995 | 89 | 513 | 7 | 14 |
In ICU Mortality (first 48h) | Admission + 48 hr | 0.044 | 0.041 | 0.289 | 0.961 | 91 | 484 | 5 | 14 |
In Hospital Mortality | Admission + 24 hr | 0.028 | 0.013 | 0.011 | 0.567 | 11 | 454 | 6 | 9 |
In Hospital Mortality | Admission + 48 hr | 0.011 | 0.060 | 0.179 | 0.964 | 84 | 631 | 7 | 13 |
LOS in ICU > 3 days | Admission + 24 hr | 0.012 | 0.090 | 0.137 | 0.626 | 26 | 650 | 8 | 14 |
LOS in ICU > 3 days | Admission + 48 hr | 0.012 | 0.049 | 0.200 | 0.960 | 84 | 615 | 7 | 13 |
LOS in Hospital > 3 days | Admission + 24 hr | 0.008 | 0.067 | 0.255 | 0.989 | 90 | 526 | 5 | 14 |
LOS in Hospital > 3 days | Admission + 48 hr | 0.001 | 0.030 | 0.028 | 0.967 | 9 | 538 | 8 | 7 |
The eICU sweep was conducted equivalently to the MIMIC-IV sweep. Please refer to the MIMIC-IV Sweep subsection above for details on the commands and sweep parameters.
For more details about eICU-specific task generation and running, see the eICU companion repository.
Task | Index Timestamp | AUC | Minimum Code Inclusion Frequency | Window Sizes | Aggregations |
---|---|---|---|---|---|
Post-discharge 30 day Mortality | Discharge | 0.603 | 68,235 | [12h,1d,full] | [code/count,value/sum_sqd,value/max] |
Post-discharge 1 year Mortality | Discharge | 0.875 | 3,280 | [30d,365d] | [static/present,value/sum,value/sum_sqd,value/min,value/max] |
In Hospital Mortality | Admission + 24 hr | 0.855 | 335,912 | [2h,7d,30d,365d,full] | [static/present,code/count,value/count,value/min,value/max] |
In Hospital Mortality | Admission + 48 hr | 0.570 | 89,121 | [12h,1d,30d] | [code/count,value/count,value/min] |
LOS in ICU > 3 days | Admission + 24 hr | 0.783 | 7,881 | [1d,30d,full] | [static/present,code/count,value/count,value/sum,value/max] |
LOS in ICU > 3 days | Admission + 48 hr | 0.757 | 1,719 | [2h,12h,7d,30d,full] | [code/count,value/count,value/sum,value/sum_sqd,value/min] |
LOS in Hospital > 3 days | Admission + 24 hr | 0.864 | 160 | [1d,30d,365d,full] | [static/present,code/count,value/min,value/max] |
LOS in Hospital > 3 days | Admission + 48 hr | 0.895 | 975 | [12h,1d,30d,365d,full] | [code/count,value/count,value/sum,value/sum_sqd] |
Task | Index Timestamp | Eta | Lambda | Alpha | Subsample | Minimum Child Weight | Number of Boosting Rounds | Early Stopping Rounds | Max Tree Depth |
---|---|---|---|---|---|---|---|---|---|
In Hospital Mortality | Admission + 24 hr | 0.043 | 0.001 | 0.343 | 0.879 | 13 | 574 | 9 | 14 |
In Hospital Mortality | Admission + 48 hr | 0.002 | 0.002 | 0.303 | 0.725 | 0 | 939 | 9 | 12 |
LOS in ICU > 3 days | Admission + 24 hr | 0.210 | 0.189 | 0.053 | 0.955 | 5 | 359 | 6 | 14 |
LOS in ICU > 3 days | Admission + 48 hr | 0.340 | 0.393 | 0.004 | 0.900 | 6 | 394 | 10 | 13 |
LOS in Hospital > 3 days | Admission + 24 hr | 0.026 | 0.238 | 0.033 | 0.940 | 46 | 909 | 5 | 11 |
LOS in Hospital > 3 days | Admission + 48 hr | 0.100 | 0.590 | 0.015 | 0.914 | 58 | 499 | 10 | 9 |
Post-discharge 30 day Mortality | Discharge | 0.003 | 0.0116 | 0.001 | 0.730 | 13 | 986 | 7 | 7 |
Post-discharge 1 year Mortality | Discharge | 0.005 | 0.006 | 0.002 | 0.690 | 93 | 938 | 6 | 14 |
Task | Index Timestamp | Number of Patients | Number of Events |
---|---|---|---|
Post-discharge 30 day Mortality | Discharge | 91,405 | 91,405 |
Post-discharge 1 year Mortality | Discharge | 91,405 | 91,405 |
In Hospital Mortality | Admission + 24 hr | 35,85 | 3,585 |
In Hospital Mortality | Admission + 48 hr | 1,527 | 1,527 |
LOS in ICU > 3 days | Admission + 24 hr | 12,672 | 14,004 |
LOS in ICU > 3 days | Admission + 48 hr | 12,712 | 14,064 |
LOS in Hospital > 3 days | Admission + 24 hr | 99,540 | 99,540 |
LOS in Hospital > 3 days | Admission + 48 hr | 99,786 | 99,786 |