CHoRUS waveform format selection

[briangow]

The requirements for the CHoRUS waveform format were outlined in: #7 . Based on this we will evaluate existing formats against these requirements. After a format is chosen (and potentially updated to meet the requirements) it will be used as the final format for the waveform data in the central CHoRUS repository. The waveform data will be distributed to the public in this format.

An initial set of potential formats were proposed in the last meeting on this topic. I will list those formats below. If you can think of any other formats which should be considered please create a new post below and I will add it to this list.

• CCDEF
• WFDB
• XML
• DICOM waveform
• EDF+
• Philips open format (.parquet based)
• Petastorm + .parquet (https://github.com/uber/petastorm)
• MWF (medical waveform format) / ISO 22077-1:2022 (https://www.iso.org/standard/82112.html)
• ROS Bags (https://wiki.ros.org/Bags)
• LabChart
• Avro
• Zarr

Once a complete list of formats has been made, we will evaluate the serious contenders for the CHoRUS waveform format against the requirements. We will be looking for volunteers to help with this evaluation.

[AEW0330]

Before anyone decides to take the time to assess the Zarr format's ability to meet each of these requirements, here is some relevant background.

It is a cloud-native format in that it supports chunking (with separate global identifiers for each chunk), compression, etc. to optimize parallel processing. But, so far as I can tell, it's mostly used for imaging and omics. I'm not aware of waveform use cases.

It was inspired by HDF5. Like HDF5, Zarr stores associated metadata in JSON and binary data in individually referenceable “chunk”-files, to provide a flexible, scalable method for storing multidimensional data.

It's purported advantages over HDF5 are:

• Zarr has better multithreading support
• Reading Zarr data with the Zarr library is significantly faster than reading HDF5 data with HDF5 libraries.
• HDF5 lacks specifications for many of the derived data types that are generated in an analysis workflow, e.g., regions of interests (ROIs), labels and other derived data created in analytic workflows. The advantages of this support in bioimaging analysis workflows are described here.
• A talk by Joe Jevnik on the differences between Zarr and HDF5

An efficiency comparison with parquet for time series - not waveform - is here).

If someone does want to go into the support of the requirements this summary of what Zarr supports may help.
Zarr supports:

• "Atomic" types: char, byte, ubyte, short, ushort, int, uint, int64, uint64.
• Shared (named) dimensions
• Attributes with specified types – both global and per-variable
• Chunking
• Fill values
• Groups
• N-Dimensional variables
• Per-variable endianness (big or little)

But since the user community is focused on its application in imaging and omics, it might not be worth a lot of further investigation unless:

• We're confident the imaging/omics binary vs. waveform binary content isn't a problem
• The purported advantages over HDF5 (e.g, faster, better threading and accommodation of labels, ROIs, annotations) and over parquet (chunking in multiple dimensions vs just 1 in parquet) are important to us.

[AEW0330]

Here is a list of assertions about problematic file format issues Bobby Greer made based on the explorations at Toronto Sick Kids done in development of ATRIUM DB.
I won't attest to their accuracy or currency, just that they seem relevant to some of the requirements listed in issue # 7.
People who can affirm or dispute these assertions could reply to this to clarify why they think these are correct or incorrect and their relevance to the requirements.

HDF5 (& CCDEF)

• has to uncompress whole file to in order to access a chunk - very inefficient

WFDB

• Often spontaneously corrupts
• Doesn't encode gaps, so have to split files - this leads to problems with managing the number of gaps typically present in files

Parquet

• requires users to predefine shard length - can't autotune
  (They are moving to support Parquet because of several advantages - principally that it is a widely used standard with less dependence on small bespoke libraries to do the majority of work with files. Other formats including the TSE format they use are niche and vulnerable to gaps in functionality and support by comparison.)

[del42]

For HDF5, compression applied on dataset/channel/chunks level. Thus, we do not need to load the whole file on memory.

[wa6gz]

If I remember correctly, the spontaneous corruption was an issue when they tried working with HDF5. I have only experienced corruptions on HDF5 file creation, not in the course of working with the files, though I typically don't edit the files heavily after creation.

[AEW0330]

If I remember correctly, the spontaneous corruption was an issue when they tried working with HDF5. I have only experienced corruptions on HDF5 file creation, not in the course of working with the files, though I typically don't edit the files heavily after creation.

Jared and I met with Bobby earlier this week. He was definitely saying this about WFDB. Again, I'm not saying anything about whether it's an accurate assertion. I'm just clarifying that he was talking about WFDB, not HDF5.

[briangow]

We have decided to eliminate the following formats from contention as the CHoRUS waveform format for the stated reasons:

• EDF+, doesn't support compression, doesn't allow for gaps (every channel has to be present in every block)
• ROS Bags, limited / clunky metadata support, lack of standardization for physiological waveforms, it's core messaging system isn't needed for our use case
• Labchart, no support for different frequencies per channel, no compression support
• Avro, no random access, no annotation support, lack of standardization for physiological waveforms
• ISO 22077, specification is not openly available

[briangow]

We have come to the conclusion that more detailed information is needed about each of the remaining formats so that we can differentiate between them and ultimately choose a format (or formats).

In particular we need more information in the following areas:

1. Benchmarking - we should do some level of benchmarking to make sure we are choosing a performant format. What benchmark tests should we run and on what data? Where should this benchmarking be run (cloud, locally)? How do we do this in a way that accurately assesses the format independent of the system / hardware that it is being run on?
2. Tooling - how extensive is the tooling and community for a given format? Is there a minimum set of tooling requirements? If so, for a format that doesn't meet those minimum requirements how much work would it be to create the needed tooling?
3. Physiological waveform specific specification - some of the remaining formats don't have physiological waveform specific specifications. For those formats we need a realistic, reasonably accurate estimate for how much time it would require to create a specification.
4. Combining formats - would combining a hierarchical data format (HDF5, Zarr) with a physiological waveform specific format (WFDB, etc.) offer advantages over either alone? If so, how much work would it be to integrate these formats with each other?
5. Choosing multiple storage formats - should we consider multiple storage formats, each targeted to handle a specific task? If so, how much additional work will it take to support more than one format? Which format (or type of format) should support which task? Will additional storage space be required to support multiple formats?

If you have thoughts on these topics, please provide them before next Thursday. If you think that combining formats (3.) or choosing multiple storage formats (4.) would be advantageous please make that known this week (and provide as much detail as possible around why that is the case), as now is the time that we need to consider this.

I will add some related notes that I took at our most recent meeting in a separate post below.

[del42]

For 1. We can use Azure ML workspace for this. Sign up with new email and get $200 credit for first month. It is sufficient amount of compute credit and we can do this in a month.
For 3. I noticed that most formats lacked metadata and annotation specs but we also not sure whether our metadata and annotations specs are going to stay within the waveform header or moved to structured db. With current requirements, can we even make realistic estimation?
For 5. It is very expensive to operate with multiple format in cloud and not advisable, especially in Azure .

[briangow]

@del42 , thanks for the thoughts!

Regarding 3., I think we need to store at least some metadata in the waveform header / file (basic information [RM1 - RM5], RM9, etc., as mentioned in #7 (comment) ). As you mention, it may be advisable to store other information in the database instead. You bring up a good point though, we may need to more clearly define what we think should be stored in the waveform header / file vs. a database sooner than later.

[briangow]

In our most recent meeting, I noted the following regarding the areas where we need more information about the remaining formats:

1. Benchmarking

• What benchmark tests should we run?
  • Input / output performance
  • Peak memory utilization
  • Time for random access of a window
  • File size / storage
  • Parallel reads / writes
  • We could use the MIMIC waveform data (perhaps p19)
• Where should the benchmarking be run?
  • In the cloud with a standardized instance configuration (consider running the benchmarking on all of the formats at one site to minimize coordination, account setup)
  • Could also consider creating test suites which could be run locally at multiple sites for each format. The benefits of doing this would be to; 1) gain insight into what sites will be facing when working with a given format; and 2) get a sense for performance across different hardware configurations. This may require a significant amount of additional work though and may not be required if the cloud testing is done.

4. Combining formats

• There was a suggestion that it may be advantageous to combine HDF5 and WFDB. The thought is that this would be particularly useful if directly accessing the PhysioNet WFDB tools would still be possible and there are benefits beyond going with one of these formats on their own. Access to the PhysioNet tools would require developing an adaptive layer. We would need to investigate how much work creating this adaptive layer would be if it is clear that there would be benefits to combining HDF5 and WFDB.

If you believe that it could be advantageous to combine HDF5 and WFDB, please make that known below before next Thursday.

5. Choosing multiple formats

• On a few occasions it has been suggested that there could be advantages to using different formats for different tasks within the CHoRUS ecosystem. @hulab-ucsf , @del42 , @sooj

If you believe that there would be advantages to using different formats for different tasks within CHoRUS, please make that known before next Thursday. Please provide as many specifics as possible (i.e. I think we should use format 1 for task A) and B) and format 2 for task C). It is advantageous to use a separate format for these tasks because...)

[agent3gatech]

Regarding 4. Combining Formats (Paraphrased from a discussion amongst Brian, Rishi and myself):

In WFDB, I think we already have a really nice standard for the data and really excellent tools in the supporting tools/packages, such as the Python WFDB package, and the other PhysioNet tools (I'm less familiar with some of the PhysioNet tools but have some experience with them).

I do worry that there might be some underestimation in the amount of work that has gone into these tools/packages and the amount of work that it would take to replicate some of these tools, to the same standard, just because of the theoretical appeal of a particular alternate file type. Either additions to the Python WFDB package to act as an 'interface layer' to a WFDB/HDF5 hybrid, or a separate interfacing package, that uses the WFDB package underneath are certainly possible, but I think should only be considered if the benefit of the hybrid format really shines over 'pure' WFDB.

If the WFDB experts/package maintainers do have updates that they would like to make to WFDB, I think that could be a good path forward, and maybe it will be worth utilizing the HDF5 files if significant updates are needed. I would really like to hear any concrete advantages utilizing HDF5 has over 'regular' WFDB (I'm not saying that there aren't any, just nobody has said).

[del42]

I may have to use AVRO for the network transfers via API calls but need to assess if APIs calls will have to consume from source data or some other processed data.

[briangow]

@del42 found this paper on Internet of Things waveform data which touches on some of the topics we are discussing: Waveform Signal Entropy and Compression Study of Whole-Building Energy Datasets . They include HDF5 and Zarr in their benchmarking analysis (primarily around compression performance) and analyze various chunk sizes. Even though their waveform data is higher resolution we may want to keep some of their methods in mind for the benchmarking that we are discussing.

[briangow]

Regarding the minimum requirements for tooling (2. above) for a given format. Our (@bemoody, @tompollard, and my) thoughts are that there should be a tool (ideally python based) which provides the ability to:

• write waveform data (samples + metadata) into the file format
• read waveform data (samples + metadata) from the file format
• write annotations into the file format
• read annotations from the file format

[del42]

@briangow When you say specification estimate? Are you referring to following? Note: it is not complete!

readChannelDataFromParquet Method Specification

Description

Reads a specific segment of waveform data from selected channels stored in a Parquet file. This method supports specifying the offset and length of the segment in either seconds or sample points, allows for optional downsampling of the data, and can return data without applying any scaling transformations. Users can specify channels to read, enhancing flexibility for data analysis.

Parameters

• filePath: str - The path to the Parquet file containing waveform data.
• offset: float - The starting point from which to read the data, interpreted in seconds or sample points based on useSecForOffset.
• length: float - The amount of data to read from the offset, interpreted in seconds or sample points based on useSecForLength.
• channels: List[int] or List[str], optional - List of channel indices or names to read. If not provided, all channels are read.
• useSecForOffset: bool - If True, interprets offset as seconds; otherwise, as sample points.
• useSecForLength: bool - If True, interprets length as seconds; otherwise, as sample points.
• downSamplingRatio: float, optional - Specifies the ratio for downsampling the data, with 1.0 indicating no downsampling. Default is 1.0.
• noDataScaling: bool, optional - If True, returns raw data without scaling. Default is False.

Returns

• dataSegment: The method returns a structured data segment (e.g., a list, NumPy array, or similar) containing the specified segment of channel data from the waveform. The data structure will include only the selected channels if specified.

Exceptions

• ValueError: Raised if invalid arguments are provided (e.g., negative lengths, offsets outside data range, invalid downsampling ratios, or invalid channel selections).
• FileNotFoundError: Raised if the specified Parquet file cannot be found.
• IOError: Raised for errors related to reading the Parquet file.

Example Usage

# Example: Reading a 5-second segment starting 10 seconds into the waveform from "waveform.parquet",
#          from channels 0 and 2, with downsampling
segment = readChannelDataFromParquet(filePath="waveform.parquet",
                                     offset=10.0,
                                     length=5.0,
                                     channels=[0, 2],
                                     useSecForOffset=True,
                                     useSecForLength=True,
                                     downSamplingRatio=2.0,
                                     noDataScaling=False)

[briangow]

In the most recent tooling meeting, it was brought up that the format we select shouldn't preclude using a second format for particular types of waveforms. For example, if EEG data is requested we may want to consider using EDF (since it a common well-supported format for EEGs) for those signals.

We do plan to include EEG data when running benchmarking for selection of the primary format for CHoRUS.

We will consider whether to include EDF as a format to be benchmarked. It was previously eliminated since it doesn't have gap or compression support.

It may be worth noting that it would be easy enough to convert from the primary CHoRUS to another format for use with specific tools, etc.

[briangow]

At the most recent waveform meeting we discussed whether to reconsider EDF as a potential option for the CHoRUS waveform format. We concluded that we would not reconsider it due to the obvious negatives (no gap or compression support).

If there are tools which we want to make use of that require EDF as the waveform format, we can convert the waveforms to EDF when using the tool.

[briangow]

At the most recent waveform meeting we discussed whether any additional benchmarking tests need to be added to #12 prior to deciding on the CHoRUS waveform format. We reviewed:

• CHoRUS waveform format selection #11 (comment)
• Waveform benchmarking #12

and concluded that the only additional test needed was a fidelity check, which will check that the read method actually returns the same data that was passed to the write method. See #12 for more details regarding the write and read methods.

We will do additional exploratory benchmarking at a later time, after the format has been chosen. This will likely include exploring more sophisticated approaches to handling gaps (by using multisegment records, etc.), and parallel reads. The group concluded that it was not likely that these other exploratory benchmarking tests would change our conclusion regarding the best format for CHoRUS. This is why we've marked them as exploratory. They will be run after the CHoRUS waveform format is chosen.

[matthewreyna]

Re: a fidelity, we are currently using the numpy.isclose function to check that the data do not exceed absolute (1e-8) and relative (1e-5) thresholds element-wise. This is helpful, but we may also want to report the maximum absolute and relative errors and/or SNR. I would not expect much variation in the errors between formats and packages if/when we start using 16-bit signed integers for all of the formats, but it may be good to quantify how much we are really returning the same data with each format, in case there is some unexpected variation. I can add it if there is interest.

[briangow]

@matthewreyna , it'd be great if you could add some additional metrics around the amount of fidelity loss. I've referenced your comment here #85. Now that we have recommended a format for CHoRUS, the topic of this discussion thread has essentially been satisfied. We can further discuss the fidelity test and any additional work around benchmarking in #85 .

[briangow]

Under the dropdown below is the characterization of a set of waveform files that I'd like to propose as a suite for running benchmarking against. The characterization provides the name, sample rate (fs), bit resolution, and length for each signal within a given waveform file.

Waveform suite characterization (click arrow to expand)

________________________________________________________________ Format: waveform_benchmark.formats.wfdb.WFDBFormat16 (WFDB with 16-bit binary storage) Record: mimic4wdb/0.1.0/waves/p100/p10079700/85594648/85594648 214981 seconds x 6 channels 255177600 timepoints, 199126720 samples (78.0%) ________________________________________________________________ Channel summary information: signal fs(Hz) Bit resolution Channel length(s) II 249.89 0.005(mV) 212497 III 249.89 0.005(mV) 5 V 249.89 0.005(mV) 212497 aVR 249.89 0.005(mV) 212492 Pleth 124.94 0.000244(NU) 212486 Resp 62.47 0.000244(Ohm) 212497 ________________________________________________________________ ________________________________________________________________ Format: waveform_benchmark.formats.wfdb.WFDBFormat16 (WFDB with 16-bit binary storage) Record: mimic4wdb/0.1.0/waves/p101/p10112163/88501826/88501826 273574 seconds x 9 channels 410181120 timepoints, 291775872 samples (71.1%) ________________________________________________________________ Channel summary information: signal fs(Hz) Bit resolution Channel length(s) II 249.89 0.005(mV) 273492 III 249.89 0.005(mV) 17654 V 249.89 0.005(mV) 273472 aVR 249.89 0.005(mV) 255821 CO2 62.47 0.0167(mmHg) 4712 CVP 124.94 0.0625(mmHg) 268406 PAP 124.94 0.0625(mmHg) 13405 Pleth 124.94 0.000244(NU) 273441 Resp 62.47 0.000244(Ohm) 273498 ________________________________________________________________ ________________________________________________________________ Format: waveform_benchmark.formats.wfdb.WFDBFormat16 (WFDB with 16-bit binary storage) Record: mimic4wdb/0.1.0/waves/p115/p11596691/86130599/86130599 172702 seconds x 5 channels 161836800 timepoints, 160764160 samples (99.3%) ________________________________________________________________ Channel summary information: signal fs(Hz) Bit resolution Channel length(s) II 249.89 0.005(mV) 172287 V 249.89 0.005(mV) 172287 aVR 249.89 0.005(mV) 172287 Pleth 124.94 0.000244(NU) 166816 Resp 62.47 0.000485(Ohm) 172282 ________________________________________________________________ ________________________________________________________________ Format: waveform_benchmark.formats.wfdb.WFDBFormat16 (WFDB with 16-bit binary storage) Record: mimic4wdb/0.1.0/waves/p136/p13620446/83988903/83988903 113279 seconds x 6 channels 134459200 timepoints, 106108032 samples (78.9%) ________________________________________________________________ Channel summary information: signal fs(Hz) Bit resolution Channel length(s) II 249.89 0.005(mV) 113274 III 249.89 0.005(mV) 5 V 249.89 0.005(mV) 113268 aVR 249.89 0.005(mV) 113263 Pleth 124.94 0.000244(NU) 112976 Resp 62.47 0.000611(Ohm) 113279 ________________________________________________________________ ________________________________________________________________ Format: waveform_benchmark.formats.wfdb.WFDBFormat16 (WFDB with 16-bit binary storage) Record: mimic4wdb/0.1.0/waves/p159/p15920699/84421559/84421559 70350 seconds x 10 channels 127452911 timepoints, 82009168 samples (64.3%) ________________________________________________________________ Channel summary information: signal fs(Hz) Bit resolution Channel length(s) I 249.89 0.005(mV) 5 II 249.89 0.005(mV) 70259 III 249.89 0.005(mV) 328 V 249.89 0.005(mV) 70263 aVR 249.89 0.005(mV) 69935 ABP 124.94 0.0625(mmHg) 66769 Ao 124.94 0.0625(mmHg) 328 CVP 124.94 0.0625(mmHg) 62261 Pleth 124.94 0.000244(NU) 70283 Resp 62.47 0.000244(Ohm) 70283 ________________________________________________________________ ________________________________________________________________ Format: waveform_benchmark.formats.wfdb.WFDBFormat16 (WFDB with 16-bit binary storage) Record: mimic4wdb/0.1.0/waves/p169/p16955095/82284982/82284982 604 seconds x 5 channels 566400 timepoints, 562240 samples (99.3%) ________________________________________________________________ Channel summary information: signal fs(Hz) Bit resolution Channel length(s) II 249.89 0.005(mV) 599 III 249.89 0.005(mV) 599 V 249.89 0.005(mV) 599 Pleth 124.94 0.000244(NU) 604 Resp 62.47 0.000611(Ohm) 599 ________________________________________________________________ ________________________________________________________________ Format: waveform_benchmark.formats.wfdb.WFDBFormat16 (WFDB with 16-bit binary storage) Record: mimic4wdb/0.1.0/waves/p178/p17802132/86917206/86917206 9650 seconds x 5 channels 9043200 timepoints, 8679040 samples (96.0%) ________________________________________________________________ Channel summary information: signal fs(Hz) Bit resolution Channel length(s) II 249.89 0.005(mV) 9261 V 249.89 0.005(mV) 9261 aVR 249.89 0.005(mV) 9261 Pleth 124.94 0.000244(NU) 9261 Resp 62.47 0.000244(Ohm) 9271 ________________________________________________________________ ________________________________________________________________ Format: waveform_benchmark.formats.wfdb.WFDBFormat16 (WFDB with 16-bit binary storage) Record: mimic4wdb/0.1.0/waves/p184/p18466737/89821554/89821554 23952 seconds x 5 channels 22444800 timepoints, 22440640 samples (100.0%) ________________________________________________________________ Channel summary information: signal fs(Hz) Bit resolution Channel length(s) II 249.89 0.005(mV) 23947 V 249.89 0.005(mV) 23947 aVR 249.89 0.005(mV) 23947 Pleth 124.94 0.000244(NU) 23952 Resp 62.47 0.000244(Ohm) 23947 ________________________________________________________________ ________________________________________________________________ Format: waveform_benchmark.formats.wfdb.WFDBFormat16 (WFDB with 16-bit binary storage) Record: mimic4wdb/0.1.0/waves/p192/p19290410/81362381/81362381 123759 seconds x 5 channels 115972800 timepoints, 46388160 samples (40.0%) ________________________________________________________________ Channel summary information: signal fs(Hz) Bit resolution Channel length(s) II 249.89 0.005(mV) 120154 V 249.89 0.005(mV) 1808 aVR 249.89 0.005(mV) 1808 Pleth 124.94 0.000244(NU) 63629 Resp 62.47 0.000928(Ohm) 120199 ________________________________________________________________ ________________________________________________________________ Format: waveform_benchmark.formats.wfdb.WFDBFormat16 (WFDB with 16-bit binary storage) Record: mimic4wdb/0.1.0/waves/p197/p19772551/83530581/83530581 89337 seconds x 6 channels 106041280 timepoints, 106041280 samples (100.0%) ________________________________________________________________ Channel summary information: signal fs(Hz) Bit resolution Channel length(s) II 249.89 0.005(mV) 89337 V2 249.89 0.005(mV) 89337 V5 249.89 0.005(mV) 89337 aVR 249.89 0.005(mV) 89337 Pleth 124.94 0.000244(NU) 89337 Resp 62.47 0.000244(Ohm) 89337 ________________________________________________________________ ________________________________________________________________ Format: waveform_benchmark.formats.wfdb.WFDBFormat16 (WFDB with 16-bit binary storage) Record: mimic3wdb/1.0/36/3654093/3654093 363240 seconds x 6 channels 272430000 timepoints, 77793875 samples (28.6%) ________________________________________________________________ Channel summary information: signal fs(Hz) Bit resolution Channel length(s) II 125.00 0.00781(mV) 327861 ABP 125.00 0.625(mmHg) 159423 PAP 125.00 0.156(mmHg) 99709 V 125.00 0.00781(mV) 35288 III 125.00 0.00781(mV) 10 I 125.00 0.00781(mV) 60 ________________________________________________________________ ________________________________________________________________ Format: waveform_benchmark.formats.wfdb.WFDBFormat16 (WFDB with 16-bit binary storage) Record: mimic3wdb/1.0/30/3033467/3033467 39540 seconds x 3 channels 14827500 timepoints, 14786000 samples (99.7%) ________________________________________________________________ Channel summary information: signal fs(Hz) Bit resolution Channel length(s) PLETH 125.00 0.00392(NU) 39507 II 125.00 0.00787(mV) 39428 V 125.00 0.00787(mV) 39353 ________________________________________________________________ ________________________________________________________________ Format: waveform_benchmark.formats.wfdb.WFDBFormat16 (WFDB with 16-bit binary storage) Record: mimic3wdb/1.0/33/3314171/3314171 101036 seconds x 3 channels 37888500 timepoints, 20224625 samples (53.4%) ________________________________________________________________ Channel summary information: signal fs(Hz) Bit resolution Channel length(s) II 125.00 0.00787(mV) 100681 I 125.00 0.00787(mV) 3 ABP 125.00 0.781(mmHg) 61113 ________________________________________________________________ ________________________________________________________________ Format: waveform_benchmark.formats.wfdb.WFDBFormat16 (WFDB with 16-bit binary storage) Record: mimic3wdb/1.0/37/3716433/3716433 198300 seconds x 6 channels 148725000 timepoints, 50687375 samples (34.1%) ________________________________________________________________ Channel summary information: signal fs(Hz) Bit resolution Channel length(s) II 125.00 0.00787(mV) 197608 V 125.00 0.00781(mV) 9406 ABP 125.00 0.781(mmHg) 189433 PLETH 125.00 0.00392(NU) 8588 RESP 125.00 0.00392(NU) 2 I 125.00 0.00392(mV) 462 ________________________________________________________________ ________________________________________________________________ Format: waveform_benchmark.formats.wfdb.WFDBFormat16 (WFDB with 16-bit binary storage) Record: mimic3wdb/1.0/35/3589404/3589404 91200 seconds x 2 channels 22800000 timepoints, 22788000 samples (99.9%) ________________________________________________________________ Channel summary information: signal fs(Hz) Bit resolution Channel length(s) II 125.00 0.00781(mV) 91152 MCL 125.00 0.00493(mV) 91152 ________________________________________________________________ ________________________________________________________________ Format: waveform_benchmark.formats.wfdb.WFDBFormat16 (WFDB with 16-bit binary storage) Record: mimic3wdb/1.0/31/3115020/3115020 113399 seconds x 2 channels 28349750 timepoints, 24644500 samples (86.9%) ________________________________________________________________ Channel summary information: signal fs(Hz) Bit resolution Channel length(s) II 125.00 0.00781(mV) 110246 ABP 125.00 0.365(mmHg) 86910 ________________________________________________________________ ________________________________________________________________ Format: waveform_benchmark.formats.wfdb.WFDBFormat16 (WFDB with 16-bit binary storage) Record: mimic3wdb/1.0/32/3279834/3279834 326520 seconds x 6 channels 244890000 timepoints, 84988625 samples (34.7%) ________________________________________________________________ Channel summary information: signal fs(Hz) Bit resolution Channel length(s) II 125.00 0.00787(mV) 322391 V 125.00 0.00787(mV) 67256 ABP 125.00 0.781(mmHg) 288037 I 125.00 0.00781(mV) 639 PLETH 125.00 0.00392(NU) 1576 RESP 125.00 0.00775(NU) 10 ________________________________________________________________ ________________________________________________________________ Format: waveform_benchmark.formats.wfdb.WFDBFormat16 (WFDB with 16-bit binary storage) Record: mimic3wdb/1.0/33/3344184/3344184 23040 seconds x 1 channels 2880000 timepoints, 2874000 samples (99.8%) ________________________________________________________________ Channel summary information: signal fs(Hz) Bit resolution Channel length(s) II 125.00 0.00787(mV) 22992 ________________________________________________________________ ________________________________________________________________ Format: waveform_benchmark.formats.wfdb.WFDBFormat16 (WFDB with 16-bit binary storage) Record: mimic3wdb/1.0/37/3740336/3740336 13080 seconds x 3 channels 4905000 timepoints, 4887375 samples (99.6%) ________________________________________________________________ Channel summary information: signal fs(Hz) Bit resolution Channel length(s) II 125.00 0.00787(mV) 13033 ABP 125.00 0.104(mmHg) 13033 PLETH 125.00 0.00392(NU) 13033 ________________________________________________________________ ________________________________________________________________ Format: waveform_benchmark.formats.wfdb.WFDBFormat16 (WFDB with 16-bit binary storage) Record: mimic3wdb/1.0/36/3663448/3663448 127500 seconds x 2 channels 31875000 timepoints, 15926375 samples (50.0%) ________________________________________________________________ Channel summary information: signal fs(Hz) Bit resolution Channel length(s) II 125.00 0.00787(mV) 127354 I 125.00 0.00781(mV) 57 ________________________________________________________________ ________________________________________________________________ Format: waveform_benchmark.formats.wfdb.WFDBFormat16 (WFDB with 16-bit binary storage) Record: charisdb/1.0.0/charis1 244797 seconds x 3 channels 36719553 timepoints, 36719553 samples (100.0%) ________________________________________________________________ Channel summary information: signal fs(Hz) Bit resolution Channel length(s) ABP 50.00 0.0116(mmHg) 244797 ECG 50.00 0.000164(mV) 244797 ICP 50.00 0.0105(mmHg) 244797 ________________________________________________________________ ________________________________________________________________ Format: waveform_benchmark.formats.wfdb.WFDBFormat16 (WFDB with 16-bit binary storage) Record: charisdb/1.0.0/charis3 341904 seconds x 3 channels 51285672 timepoints, 51285672 samples (100.0%) ________________________________________________________________ Channel summary information: signal fs(Hz) Bit resolution Channel length(s) ABP 50.00 0.0124(mmHg) 341904 ECG 50.00 0.000164(mV) 341904 ICP 50.00 0.0125(mmHg) 341904 ________________________________________________________________ ________________________________________________________________ Format: waveform_benchmark.formats.wfdb.WFDBFormat16 (WFDB with 16-bit binary storage) Record: charisdb/1.0.0/charis4 144000 seconds x 3 channels 21599997 timepoints, 21599997 samples (100.0%) ________________________________________________________________ Channel summary information: signal fs(Hz) Bit resolution Channel length(s) ABP 50.00 0.0109(mmHg) 144000 ECG 50.00 0.000164(mV) 144000 ICP 50.00 0.0119(mmHg) 144000 ________________________________________________________________ ________________________________________________________________ Format: waveform_benchmark.formats.wfdb.WFDBFormat16 (WFDB with 16-bit binary storage) Record: charisdb/1.0.0/charis5 719994 seconds x 3 channels 107999165 timepoints, 107999166 samples (100.0%) ________________________________________________________________ Channel summary information: signal fs(Hz) Bit resolution Channel length(s) ABP 50.00 0.0124(mmHg) 719994 ECG 50.00 0.000164(mV) 719994 ICP 50.00 0.012(mmHg) 719994 ________________________________________________________________ ________________________________________________________________ Format: waveform_benchmark.formats.wfdb.WFDBFormat16 (WFDB with 16-bit binary storage) Record: charisdb/1.0.0/charis7 483615 seconds x 3 channels 72542202 timepoints, 72542202 samples (100.0%) ________________________________________________________________ Channel summary information: signal fs(Hz) Bit resolution Channel length(s) ABP 50.00 0.0125(mmHg) 483615 ECG 50.00 0.000164(mV) 483615 ICP 50.00 0.0102(mmHg) 483615 ________________________________________________________________ ________________________________________________________________ Format: waveform_benchmark.formats.wfdb.WFDBFormat16 (WFDB with 16-bit binary storage) Record: charisdb/1.0.0/charis8 343399 seconds x 3 channels 51509871 timepoints, 51509871 samples (100.0%) ________________________________________________________________ Channel summary information: signal fs(Hz) Bit resolution Channel length(s) ABP 50.00 0.0113(mmHg) 343399 ECG 50.00 0.000164(mV) 343399 ICP 50.00 0.012(mmHg) 343399 ________________________________________________________________ ________________________________________________________________ Format: waveform_benchmark.formats.wfdb.WFDBFormat16 (WFDB with 16-bit binary storage) Record: charisdb/1.0.0/charis9 615600 seconds x 3 channels 92339937 timepoints, 92339937 samples (100.0%) ________________________________________________________________ Channel summary information: signal fs(Hz) Bit resolution Channel length(s) ABP 50.00 0.0124(mmHg) 615600 ECG 50.00 0.000164(mV) 615600 ICP 50.00 0.00774(mmHg) 615600 ________________________________________________________________ ________________________________________________________________ Format: waveform_benchmark.formats.wfdb.WFDBFormat16 (WFDB with 16-bit binary storage) Record: charisdb/1.0.0/charis10 425575 seconds x 3 channels 63836319 timepoints, 63836319 samples (100.0%) ________________________________________________________________ Channel summary information: signal fs(Hz) Bit resolution Channel length(s) ABP 50.00 0.0125(mmHg) 425575 ECG 50.00 0.000164(mV) 425575 ICP 50.00 0.0125(mmHg) 425575 ________________________________________________________________ ________________________________________________________________ Format: waveform_benchmark.formats.wfdb.WFDBFormat16 (WFDB with 16-bit binary storage) Record: charisdb/1.0.0/charis11 346568 seconds x 3 channels 51985170 timepoints, 51985170 samples (100.0%) ________________________________________________________________ Channel summary information: signal fs(Hz) Bit resolution Channel length(s) ABP 50.00 0.0124(mmHg) 346568 ECG 50.00 0.000164(mV) 346568 ICP 50.00 0.0116(mmHg) 346568 ________________________________________________________________ ________________________________________________________________ Format: waveform_benchmark.formats.wfdb.WFDBFormat16 (WFDB with 16-bit binary storage) Record: charisdb/1.0.0/charis12 410903 seconds x 3 channels 61635398 timepoints, 61635399 samples (100.0%) ________________________________________________________________ Channel summary information: signal fs(Hz) Bit resolution Channel length(s) ABP 50.00 0.0164(mmHg) 410903 ECG 50.00 0.000164(mV) 410903 ICP 50.00 0.0164(mmHg) 410903 ________________________________________________________________ ________________________________________________________________ Format: waveform_benchmark.formats.wfdb.WFDBFormat16 (WFDB with 16-bit binary storage) Record: challenge-2018/1.0.0/training/tr03-0079/tr03-0079 26815 seconds x 13 channels 69719000 timepoints, 69719000 samples (100.0%) ________________________________________________________________ Channel summary information: signal fs(Hz) Bit resolution Channel length(s) F3-M2 200.00 1(uV) 26815 F4-M1 200.00 1(uV) 26815 C3-M2 200.00 1(uV) 26815 C4-M1 200.00 1(uV) 26815 O1-M2 200.00 1(uV) 26815 O2-M1 200.00 1(uV) 26815 E1-M2 200.00 1(uV) 26815 Chin1-Chin2 200.00 1(uV) 26815 ABD 200.00 1(uV) 26815 CHEST 200.00 1(uV) 26815 AIRFLOW 200.00 1(uV) 26815 SaO2 200.00 0.00153(%) 26815 ECG 200.00 0.001(mV) 26815 ________________________________________________________________ ________________________________________________________________ Format: waveform_benchmark.formats.wfdb.WFDBFormat16 (WFDB with 16-bit binary storage) Record: challenge-2018/1.0.0/training/tr03-0146/tr03-0146 27895 seconds x 13 channels 72527000 timepoints, 72527000 samples (100.0%) ________________________________________________________________ Channel summary information: signal fs(Hz) Bit resolution Channel length(s) F3-M2 200.00 1(uV) 27895 F4-M1 200.00 1(uV) 27895 C3-M2 200.00 1(uV) 27895 C4-M1 200.00 1(uV) 27895 O1-M2 200.00 1(uV) 27895 O2-M1 200.00 1(uV) 27895 E1-M2 200.00 1(uV) 27895 Chin1-Chin2 200.00 1(uV) 27895 ABD 200.00 1(uV) 27895 CHEST 200.00 1(uV) 27895 AIRFLOW 200.00 1(uV) 27895 SaO2 200.00 0.00153(%) 27895 ECG 200.00 0.001(mV) 27895 ________________________________________________________________ ________________________________________________________________ Format: waveform_benchmark.formats.wfdb.WFDBFormat16 (WFDB with 16-bit binary storage) Record: challenge-2018/1.0.0/training/tr05-1675/tr05-1675 30460 seconds x 13 channels 79196000 timepoints, 79196000 samples (100.0%) ________________________________________________________________ Channel summary information: signal fs(Hz) Bit resolution Channel length(s) F3-M2 200.00 1(uV) 30460 F4-M1 200.00 1(uV) 30460 C3-M2 200.00 1(uV) 30460 C4-M1 200.00 1(uV) 30460 O1-M2 200.00 1(uV) 30460 O2-M1 200.00 1(uV) 30460 E1-M2 200.00 1(uV) 30460 Chin1-Chin2 200.00 1(uV) 30460 ABD 200.00 1(uV) 30460 CHEST 200.00 1(uV) 30460 AIRFLOW 200.00 1(uV) 30460 SaO2 200.00 0.00153(%) 30460 ECG 200.00 0.001(mV) 30460 ________________________________________________________________ ________________________________________________________________ Format: waveform_benchmark.formats.wfdb.WFDBFormat16 (WFDB with 16-bit binary storage) Record: challenge-2018/1.0.0/training/tr06-0107/tr06-0107 24355 seconds x 13 channels 63323000 timepoints, 63323000 samples (100.0%) ________________________________________________________________ Channel summary information: signal fs(Hz) Bit resolution Channel length(s) F3-M2 200.00 1(uV) 24355 F4-M1 200.00 1(uV) 24355 C3-M2 200.00 1(uV) 24355 C4-M1 200.00 1(uV) 24355 O1-M2 200.00 1(uV) 24355 O2-M1 200.00 1(uV) 24355 E1-M2 200.00 1(uV) 24355 Chin1-Chin2 200.00 1(uV) 24355 ABD 200.00 1(uV) 24355 CHEST 200.00 1(uV) 24355 AIRFLOW 200.00 1(uV) 24355 SaO2 200.00 0.00153(%) 24355 ECG 200.00 0.001(mV) 24355 ________________________________________________________________ ________________________________________________________________ Format: waveform_benchmark.formats.wfdb.WFDBFormat16 (WFDB with 16-bit binary storage) Record: challenge-2018/1.0.0/training/tr08-0351/tr08-0351 28812 seconds x 13 channels 74911200 timepoints, 74911200 samples (100.0%) ________________________________________________________________ Channel summary information: signal fs(Hz) Bit resolution Channel length(s) F3-M2 200.00 1(uV) 28812 F4-M1 200.00 1(uV) 28812 C3-M2 200.00 1(uV) 28812 C4-M1 200.00 1(uV) 28812 O1-M2 200.00 1(uV) 28812 O2-M1 200.00 1(uV) 28812 E1-M2 200.00 1(uV) 28812 Chin1-Chin2 200.00 1(uV) 28812 ABD 200.00 1(uV) 28812 CHEST 200.00 1(uV) 28812 AIRFLOW 200.00 1(uV) 28812 SaO2 200.00 0.00153(%) 28812 ECG 200.00 0.001(mV) 28812 ________________________________________________________________ ________________________________________________________________ Format: waveform_benchmark.formats.wfdb.WFDBFormat16 (WFDB with 16-bit binary storage) Record: challenge-2018/1.0.0/training/tr09-0593/tr09-0593 29292 seconds x 13 channels 76159200 timepoints, 76159200 samples (100.0%) ________________________________________________________________ Channel summary information: signal fs(Hz) Bit resolution Channel length(s) F3-M2 200.00 1(uV) 29292 F4-M1 200.00 1(uV) 29292 C3-M2 200.00 1(uV) 29292 C4-M1 200.00 1(uV) 29292 O1-M2 200.00 1(uV) 29292 O2-M1 200.00 1(uV) 29292 E1-M2 200.00 1(uV) 29292 Chin1-Chin2 200.00 1(uV) 29292 ABD 200.00 1(uV) 29292 CHEST 200.00 1(uV) 29292 AIRFLOW 200.00 1(uV) 29292 SaO2 200.00 0.00153(%) 29292 ECG 200.00 0.001(mV) 29292 ________________________________________________________________ ________________________________________________________________ Format: waveform_benchmark.formats.wfdb.WFDBFormat16 (WFDB with 16-bit binary storage) Record: challenge-2018/1.0.0/training/tr10-0423/tr10-0423 29408 seconds x 13 channels 76460800 timepoints, 76460800 samples (100.0%) ________________________________________________________________ Channel summary information: signal fs(Hz) Bit resolution Channel length(s) F3-M2 200.00 1(uV) 29408 F4-M1 200.00 1(uV) 29408 C3-M2 200.00 1(uV) 29408 C4-M1 200.00 1(uV) 29408 O1-M2 200.00 1(uV) 29408 O2-M1 200.00 1(uV) 29408 E1-M2 200.00 1(uV) 29408 Chin1-Chin2 200.00 1(uV) 29408 ABD 200.00 1(uV) 29408 CHEST 200.00 1(uV) 29408 AIRFLOW 200.00 1(uV) 29408 SaO2 200.00 0.00153(%) 29408 ECG 200.00 0.001(mV) 29408 ________________________________________________________________ ________________________________________________________________ Format: waveform_benchmark.formats.wfdb.WFDBFormat16 (WFDB with 16-bit binary storage) Record: challenge-2018/1.0.0/training/tr11-0042/tr11-0042 29296 seconds x 13 channels 76169600 timepoints, 76169600 samples (100.0%) ________________________________________________________________ Channel summary information: signal fs(Hz) Bit resolution Channel length(s) F3-M2 200.00 1(uV) 29296 F4-M1 200.00 1(uV) 29296 C3-M2 200.00 1(uV) 29296 C4-M1 200.00 1(uV) 29296 O1-M2 200.00 1(uV) 29296 O2-M1 200.00 1(uV) 29296 E1-M2 200.00 1(uV) 29296 Chin1-Chin2 200.00 1(uV) 29296 ABD 200.00 1(uV) 29296 CHEST 200.00 1(uV) 29296 AIRFLOW 200.00 1(uV) 29296 SaO2 200.00 0.00153(%) 29296 ECG 200.00 0.001(mV) 29296 ________________________________________________________________ ________________________________________________________________ Format: waveform_benchmark.formats.wfdb.WFDBFormat16 (WFDB with 16-bit binary storage) Record: challenge-2018/1.0.0/training/tr12-0185/tr12-0185 26692 seconds x 13 channels 69399200 timepoints, 69399200 samples (100.0%) ________________________________________________________________ Channel summary information: signal fs(Hz) Bit resolution Channel length(s) F3-M2 200.00 1(uV) 26692 F4-M1 200.00 1(uV) 26692 C3-M2 200.00 1(uV) 26692 C4-M1 200.00 1(uV) 26692 O1-M2 200.00 1(uV) 26692 O2-M1 200.00 1(uV) 26692 E1-M2 200.00 1(uV) 26692 Chin1-Chin2 200.00 1(uV) 26692 ABD 200.00 1(uV) 26692 CHEST 200.00 1(uV) 26692 AIRFLOW 200.00 1(uV) 26692 SaO2 200.00 0.00153(%) 26692 ECG 200.00 0.001(mV) 26692 ________________________________________________________________ ________________________________________________________________ Format: waveform_benchmark.formats.wfdb.WFDBFormat16 (WFDB with 16-bit binary storage) Record: challenge-2018/1.0.0/training/tr14-0272/tr14-0272 30152 seconds x 13 channels 78395200 timepoints, 78395200 samples (100.0%) ________________________________________________________________ Channel summary information: signal fs(Hz) Bit resolution Channel length(s) F3-M2 200.00 1(uV) 30152 F4-M1 200.00 1(uV) 30152 C3-M2 200.00 1(uV) 30152 C4-M1 200.00 1(uV) 30152 O1-M2 200.00 1(uV) 30152 O2-M1 200.00 1(uV) 30152 E1-M2 200.00 1(uV) 30152 Chin1-Chin2 200.00 1(uV) 30152 ABD 200.00 1(uV) 30152 CHEST 200.00 1(uV) 30152 AIRFLOW 200.00 1(uV) 30152 SaO2 200.00 0.00153(%) 30152 ECG 200.00 0.001(mV) 30152 ________________________________________________________________

This suite consists of 10 files from each of these databases (on PhysioNet):

1. MIMIC-IV Waveform Database, version 0.1.0: ICU monitor data from Philips - DWC
2. MIMIC-III Waveform Database, version 1.0: ICU monitor data from Philips - RDE
3. CHARIS database, version 1.0.0: SICU monitor data from GE (with an ICP signal in each file)
4. You Snooze You Win: The PhysioNet/Computing in Cardiology Challenge 2018, version 1.0.0: Sleep lab data (EEG, EOG, EMG, etc.)

The length of these files varies from minutes to 8+ days.

We may also add an additional 10 waveform files of ICU monitor data from GE. We will post details regarding those files below once they are available.

Finally, we plan to test each format to make sure it can read and write a very long file. For this we could use this file from the MIMIC-III Waveform Database which spans across 25 days:

________________________________________________________________
Format: waveform_benchmark.formats.wfdb.WFDBFormat16
         (WFDB with 16-bit binary storage)
Record: /home/briangow/chorus/data/waveform_standardization/benchmarking/mimic_iii_wdb/34/3453132/3453132
         2160123 seconds x 3 channels
         810046125 timepoints, 718116750 samples (88.7%)
________________________________________________________________
Channel summary information:
 signal       fs(Hz)     Bit resolution       Channel length(s)   
 II           125.00     0.0123(mV)           2151219             
 PLETH        125.00     0.00392(NU)          2141078             
 ABP          125.00     0.781(mmHg)          1452637             
________________________________________________________________

Please share your thoughts below regarding this plan.

[briangow]

These files from a GE ICU monitor are also being run as part of the benchmarking:

GE ICU monitor waveform characterization (click arrow to expand)

________________________________________________________________ Format: waveform_benchmark.formats.wfdb.WFDBFormat16 (WFDB with 16-bit binary storage) Record: /home/briangow/chorus/data/waveform_benchmark/waveforms/GE_WFDB/A004-0502133511/A004-0502133511 436270 seconds x 8 channels 837638400 timepoints, 818860800 samples (97.8%) ________________________________________________________________ Channel summary information: signal fs(Hz) Bit resolution Channel length(s) I 240.00 0.00244(mV) 433690 II 240.00 0.00244(mV) 433690 III 240.00 0.00244(mV) 433690 V 240.00 0.00244(mV) 433690 RR 240.00 0.005(mV) 433690 SPO2 240.00 0.005(mV) 433690 AR1 240.00 0.2(mmHg) 404890 CVP2 240.00 0.2(mmHg) 404890 ________________________________________________________________ ________________________________________________________________ Format: waveform_benchmark.formats.wfdb.WFDBFormat16 (WFDB with 16-bit binary storage) Record: /home/briangow/chorus/data/waveform_benchmark/waveforms/GE_WFDB/A004-0502133511/A004-0502133511 436270 seconds x 8 channels 837638400 timepoints, 818860800 samples (97.8%) ________________________________________________________________ Channel summary information: signal fs(Hz) Bit resolution Channel length(s) I 240.00 0.00244(mV) 433690 II 240.00 0.00244(mV) 433690 III 240.00 0.00244(mV) 433690 V 240.00 0.00244(mV) 433690 RR 240.00 0.005(mV) 433690 SPO2 240.00 0.005(mV) 433690 AR1 240.00 0.2(mmHg) 404890 CVP2 240.00 0.2(mmHg) 404890 ________________________________________________________________ ________________________________________________________________ Format: waveform_benchmark.formats.wfdb.WFDBFormat16 (WFDB with 16-bit binary storage) Record: /home/briangow/chorus/data/waveform_benchmark/waveforms/GE_WFDB/A004-0502569551/A004-0502569551 23528 seconds x 7 channels 39527040 timepoints, 39527040 samples (100.0%) ________________________________________________________________ Channel summary information: signal fs(Hz) Bit resolution Channel length(s) I 240.00 0.00244(mV) 23528 II 240.00 0.00244(mV) 23528 III 240.00 0.00244(mV) 23528 V 240.00 0.00244(mV) 23528 AR1 240.00 0.2(mmHg) 23528 RR 240.00 0.005(mV) 23528 SPO2 240.00 0.005(mV) 23528 ________________________________________________________________ ________________________________________________________________ Format: waveform_benchmark.formats.wfdb.WFDBFormat16 (WFDB with 16-bit binary storage) Record: /home/briangow/chorus/data/waveform_benchmark/waveforms/GE_WFDB/A004-0502765334/A004-0502765334 143832 seconds x 6 channels 207118080 timepoints, 207118080 samples (100.0%) ________________________________________________________________ Channel summary information: signal fs(Hz) Bit resolution Channel length(s) I 240.00 0.00244(mV) 143832 II 240.00 0.00244(mV) 143832 III 240.00 0.00244(mV) 143832 V 240.00 0.00244(mV) 143832 RR 240.00 0.005(mV) 143832 SPO2 240.00 0.005(mV) 143832 ________________________________________________________________ ________________________________________________________________ Format: waveform_benchmark.formats.wfdb.WFDBFormat16 (WFDB with 16-bit binary storage) Record: /home/briangow/chorus/data/waveform_benchmark/waveforms/GE_WFDB/A004-0500372071/A004-0500372071 202020 seconds x 9 channels 436363200 timepoints, 436363200 samples (100.0%) ________________________________________________________________ Channel summary information: signal fs(Hz) Bit resolution Channel length(s) I 240.00 0.00244(mV) 202020 II 240.00 0.00244(mV) 202020 III 240.00 0.00244(mV) 202020 V 240.00 0.00244(mV) 202020 AR1 240.00 0.2(mmHg) 202020 PA2 240.00 0.2(mmHg) 202020 SPO2 240.00 0.005(mV) 202020 RR 240.00 0.005(mV) 202020 CVP2 240.00 0.2(mmHg) 202020 ________________________________________________________________ ________________________________________________________________ Format: waveform_benchmark.formats.wfdb.WFDBFormat16 (WFDB with 16-bit binary storage) Record: /home/briangow/chorus/data/waveform_benchmark/waveforms/GE_WFDB/A004-0500588750/A004-0500588750 341914 seconds x 7 channels 574415520 timepoints, 515346720 samples (89.7%) ________________________________________________________________ Channel summary information: signal fs(Hz) Bit resolution Channel length(s) I 240.00 0.00244(mV) 306754 II 240.00 0.00244(mV) 306754 III 240.00 0.00244(mV) 306754 V 240.00 0.00244(mV) 306754 SPO2 240.00 0.005(mV) 306754 RR 240.00 0.005(mV) 306754 AR1 240.00 0.2(mmHg) 306754 ________________________________________________________________ ________________________________________________________________ Format: waveform_benchmark.formats.wfdb.WFDBFormat16 (WFDB with 16-bit binary storage) Record: /home/briangow/chorus/data/waveform_benchmark/waveforms/GE_WFDB/A004-0503168997/A004-0503168997 514372 seconds x 7 channels 864144960 timepoints, 822672960 samples (95.2%) ________________________________________________________________ Channel summary information: signal fs(Hz) Bit resolution Channel length(s) I 240.00 0.00244(mV) 514372 II 240.00 0.00244(mV) 514372 III 240.00 0.00244(mV) 514372 V 240.00 0.00244(mV) 514372 SPO2 240.00 0.005(mV) 514372 RR 240.00 0.005(mV) 514372 AR1 240.00 0.2(mmHg) 341572 ________________________________________________________________ ________________________________________________________________ Format: waveform_benchmark.formats.wfdb.WFDBFormat16 (WFDB with 16-bit binary storage) Record: /home/briangow/chorus/data/waveform_benchmark/waveforms/GE_WFDB/A004-0502617045/A004-0502617045 132450 seconds x 6 channels 190728000 timepoints, 190728000 samples (100.0%) ________________________________________________________________ Channel summary information: signal fs(Hz) Bit resolution Channel length(s) I 240.00 0.00244(mV) 132450 II 240.00 0.00244(mV) 132450 III 240.00 0.00244(mV) 132450 V 240.00 0.00244(mV) 132450 SPO2 240.00 0.005(mV) 132450 RR 240.00 0.005(mV) 132450 ________________________________________________________________ ________________________________________________________________ Format: waveform_benchmark.formats.wfdb.WFDBFormat16 (WFDB with 16-bit binary storage) Record: /home/briangow/chorus/data/waveform_benchmark/waveforms/GE_WFDB/A004-0500184533/A004-0500184533 73982 seconds x 6 channels 106534080 timepoints, 101782080 samples (95.5%) ________________________________________________________________ Channel summary information: signal fs(Hz) Bit resolution Channel length(s) I 240.00 0.00244(mV) 70682 II 240.00 0.00244(mV) 70682 III 240.00 0.00244(mV) 70682 V 240.00 0.00244(mV) 70682 SPO2 240.00 0.005(mV) 70682 RR 240.00 0.005(mV) 70682 ________________________________________________________________ ________________________________________________________________ Format: waveform_benchmark.formats.wfdb.WFDBFormat16 (WFDB with 16-bit binary storage) Record: /home/briangow/chorus/data/waveform_benchmark/waveforms/GE_WFDB/A004-0502940243/A004-0502940243 156130 seconds x 7 channels 262298400 timepoints, 262298400 samples (100.0%) ________________________________________________________________ Channel summary information: signal fs(Hz) Bit resolution Channel length(s) I 240.00 0.00244(mV) 156130 II 240.00 0.00244(mV) 156130 III 240.00 0.00244(mV) 156130 V 240.00 0.00244(mV) 156130 RR 240.00 0.005(mV) 156130 SPO2 240.00 0.005(mV) 156130 CO2 240.00 0.25(mmHg) 156130 ________________________________________________________________ ________________________________________________________________ Format: waveform_benchmark.formats.wfdb.WFDBFormat16 (WFDB with 16-bit binary storage) Record: /home/briangow/chorus/data/waveform_benchmark/waveforms/GE_WFDB/A004-0502604815/A004-0502604815 578 seconds x 6 channels 832320 timepoints, 832320 samples (100.0%) ________________________________________________________________ Channel summary information: signal fs(Hz) Bit resolution Channel length(s) I 240.00 0.00244(mV) 578 II 240.00 0.00244(mV) 578 III 240.00 0.00244(mV) 578 V 240.00 0.00244(mV) 578 RR 240.00 0.005(mV) 578 SPO2 240.00 0.005(mV) 578 ________________________________________________________________

[briangow]

The p102 waveform from MIMIC-IV (with 5s ECG# channels) is being replaced with this waveform:

Record: 
/data/waveform_benchmark/waveforms/mimic4wdb/0.1.0/waves/p105/p10560354/81105139/81105139
         20279 seconds x 3 channels
         8868160 timepoints, 8866560 samples (100.0%)
________________________________________________________________
Channel summary information:
 signal       fs(Hz)     Bit resolution       Channel length(s)   
 II           249.89     0.005(mV)            20274               
 Pleth        124.94     0.000244(NU)         20279               
 Resp         62.47      0.000928(Ohm)        20274               
________________________________________________________________ 

[briangow]

Numpy, VitalDB (https://vitaldb.net/docs/), and AtriumDB (https://laussenlabs.ca/atriumdb/) formats are also being evaluated.

[briangow]

Here are the results of the benchmarking run used as the basis for making plots for our next meeting.

benchmark_results.log

[briangow]

At our most recent meeting, the working group suggested that WFDB should be used as the basis for a waveform format for CHoRUS. While some other formats, such as CCDEF and Zarr, had advantages, their lack of complete physiological waveform specification led to the group recommending WFDB. Like some other formats, WFDB also has the benefit of being interoperable, is a mature format with solid software and tooling, performed well on the benchmarking tests, and has a large user community.

There was some interest in combining HDF5 and WFDB but we concluded that this is something we could scope out and do in a subsequent project.

We plan to share some of the work that the working group did with other groups (across Bridge2AI) to solicit their feedback.

[briangow]

In addition to adding more of the tests proposed above to the benchmarking code that we ran as part of this process, we may want to consider adding additional constrictions. In particular, we should think about forcing consistency across the formats in these areas:

• by using the same data type (ex: 16 bit signed integers) with ADC gain and offset
• by using same approach for storing arrays with respect to whether they are contiguous in memory (i.e. storing as row-major, column-major, or intentional)

[briangow]

The working group has selected WFDB as the base format for waveforms in CHoRUS. I'm going to close this discussion since the topic of this thread has been satisfied - related discussions will continue elsewhere.