WEBVTT 00:00.200 --> 00:08.640 Let's explore how to use guardrails AI framework to systematically process and validate medical data 00:08.640 --> 00:11.240 records derived from doctor's notes. 00:14.840 --> 00:18.560 By employing guardrails, AI's a built in validators. 00:19.480 --> 00:27.600 The application underscores the importance of structured data validation in healthcare information processing. 00:30.000 --> 00:34.560 The program starts by importing necessary modules and packages. 00:35.800 --> 00:43.320 Base model and field from parenthesis are used to define data models with specific attributes and validation 00:43.320 --> 00:43.960 rules. 00:46.440 --> 00:55.760 The guardrails framework is represented by GD, which facilitates the validation process alongside valid 00:55.760 --> 00:59.640 range and valid choices. 01:00.230 --> 01:07.230 Additional imports include OpenAI for accessing AI driven text processing capabilities. 01:09.110 --> 01:14.190 JSON for data formatting and reach for enhanced console printing. 01:18.150 --> 01:28.550 The next step would be to define the output model and prompt and build rail specification using these 01:28.550 --> 01:32.390 two components and creating a guard object. 01:33.670 --> 01:44.430 So in this case, the output models are defined using symptom medication and patient info models. 01:46.030 --> 01:47.950 The symptom model captures. 01:50.670 --> 01:54.350 Medical symptoms experienced by a patient. 01:55.430 --> 02:03.370 Each symptom is associated with an effective body part with validation ensuring only specified areas 02:03.370 --> 02:07.170 such as head, neck and chest are accepted. 02:08.250 --> 02:12.490 This precision is crucial for accurate medical record keeping. 02:16.050 --> 02:25.690 The medication model documents medications a patient is taking, including medication name and patient's 02:25.690 --> 02:26.610 response. 02:27.250 --> 02:34.690 This model is essential for tracking treatment efficiency and patient reactions. 02:35.050 --> 02:44.170 Finally, the patient model aggregates comprehensive patient information including gender, age, a 02:44.170 --> 02:46.770 list of symptoms, and current medications. 02:51.610 --> 02:57.930 Age validation checks are realistic values between 0 to 100 years. 02:58.090 --> 03:00.720 Enforcing logical data integrity. 03:04.240 --> 03:06.160 List symptoms and medication. 03:06.280 --> 03:08.440 Tie back to their respective models. 03:08.760 --> 03:12.320 Ensuring a detailed, validated patient profile. 03:15.040 --> 03:21.200 Now that we have defined the output model, let's construct the AI prompt. 03:26.160 --> 03:33.240 A meticulously crafted prompt instructs the AI to analyze the doctor's notes and extract structured 03:33.240 --> 03:39.000 patient information, aligning with defined pydantic models. 03:41.360 --> 03:48.840 This approach leverages AI's potential to interpret complex medical text, transforming unstructured 03:48.840 --> 03:56.160 data into a structured format for further analysis. 03:57.440 --> 04:05.510 In this case, The doctor's notes will be replaced by the value defined in this variable, which states 04:05.510 --> 04:14.430 45 year old male with chronic macular rash to face and hair worse in beard, eyebrows and nares. 04:14.470 --> 04:17.150 Itchy, flaky, slightly scaly. 04:17.430 --> 04:23.230 Moderate response to OTC steroid system. 04:23.230 --> 04:32.550 And finally, this is built in variable within the Gardel serie framework that would define a structured 04:32.550 --> 04:41.150 JSON example that can help the LLM to create a valid JSON structure. 04:41.150 --> 04:51.030 Once the prompts and output models are defined, this presents the base for creating the guard object. 04:52.350 --> 04:58.310 The guard object is instantiated with patient info model, setting the stage for Garters framework to 04:58.350 --> 05:02.210 enforce data validation based on the model's constraints. 05:04.250 --> 05:12.810 This integration is pivotal for translating AI generated insights into validated structures that adhere 05:12.810 --> 05:19.050 to predefined medical and data integrity standards. 05:19.090 --> 05:19.730 Finally. 05:21.850 --> 05:35.770 The call to an LLM is wrapped into the guard object, which basically invokes a call to LLM and validates 05:35.770 --> 05:38.570 the output and performs necessary action. 05:41.370 --> 05:43.170 In this case, we have defined. 05:45.850 --> 05:55.770 That if the affected area of symptom is outside these three varied choices, the LLM should be re-asked 05:55.770 --> 05:57.490 to generate the output. 05:59.120 --> 06:05.200 This action is performed internally by the guardrail CF framework within this call. 06:08.640 --> 06:16.200 Let's execute this program and understand how it works. 06:38.640 --> 06:39.880 As you can see here. 06:42.280 --> 06:50.840 This is the initial step in which the prompt prepared by guardrails AI framework is given as an input 06:50.840 --> 06:51.760 to LLM. 06:52.760 --> 07:00.430 This includes doctors notes and the output model that we are expecting. 07:02.470 --> 07:06.630 Which includes patients gender age in the varied range. 07:07.870 --> 07:10.110 Symptom. 07:10.110 --> 07:12.790 Valid choices current medications. 07:15.710 --> 07:19.150 And then this is the raw LLM output generated by. 07:19.310 --> 07:25.310 LLM which includes gender symptoms and other information. 07:26.270 --> 07:29.190 This output is validated by cartels. 07:29.230 --> 07:35.910 AI framework against the criteria that has been defined in this case. 07:36.350 --> 07:37.470 As you can see here. 07:41.790 --> 07:45.510 Some of the values doesn't match our valid choices. 07:45.510 --> 07:47.350 So value face and head. 07:47.430 --> 07:55.350 Is not in the choice for the affected area and hence guardrail. 07:55.470 --> 08:08.010 Say I prepare Another call against LM by modifying the prompt, which says some of the input doesn't 08:08.010 --> 08:09.970 match to my criteria. 08:10.970 --> 08:13.170 Can you regenerate the output? 08:13.290 --> 08:17.210 And then LM again creates another output model. 08:17.330 --> 08:23.010 As you can see here which matches our defined standards. 08:23.570 --> 08:28.330 And this output is again validated by the guardrail CI framework. 08:28.690 --> 08:30.690 And then can be presented here. 08:40.730 --> 08:49.130 Since this program demonstrates a holistic approach to health care data processing, leveraging AI capabilities 08:49.130 --> 08:57.530 and stringent data validation to ensure accuracy, compliance and integrity in medical record keeping.