Chapter 235: De-Identification and Confidence

One forty in the morning. The office air conditioner had long since shut off, and the air was heavy with a damp, muggy heat. Lin Chen had his left foot propped on a spare computer tower beside him, two thick dictionaries wedged under his knee. The sharp pain in his ankle had dulled into a heavy, throbbing ache, like a thin iron wire slowly twisting between the bones. He didn’t rub it. He only screwed the cap back onto the pain-relief balm and returned it to the drawer. Inside the drawer there was still half a blister pack of ibuprofen and a crumpled payment slip, its edges already yellowing, gray ground into the folds so deep it would never wash out.

On the screen were the raw forms for fifty de-identified medical records. The template from the procurement office was very clear: patient basic information, chief complaint, past medical history, laboratory tests, imaging findings, model inference results, confidence score, and degree of agreement with the attending physician’s final diagnosis. There weren’t many fields, but the raw data exported from the hospital looked like a pot of porridge that had never been properly stirred. Date formats mixed “2023-05-12” with “2023/5/12.” Some lab indicators came with units, some were just numbers. The chief complaint fields were stuffed with colloquial dialect expressions and typos. Worse, some fields had been automatically truncated by the hospital’s old HIS system, ending in garbled characters like the torn edge of paper cut with a dull blade.

“Case thirty-seven—creatinine value missing.” Su Man’s voice came from the neighboring desk, rough with the rasp of staying up all night. She didn’t look up, just slid her fingers quickly over the trackpad. “The model only gave it a confidence score of 0.41. Should we fill it with the median? Or use KNN imputation?”

“Don’t fill it.” Lin Chen stared at the screen and typed a few more lines of code. “Mark missing values in red. Add a note to the inference result: ‘requires manual review.’ What the hospital wants is real-world validation, not a perfect answer sheet. If we force the data to fit, any compliance audit that checks the logs will see right through it. Clinical data isn’t a math problem. If it’s missing, it’s missing. A fabricated number can fool a report, but it won’t fool a doctor.”

Su Man stopped moving for a moment and turned to look at him. The cold light from the screen washed over her face, and the shadows under her eyes were dark and deep. “Then the agreement rate will drop below seventy-eight percent. The procurement office’s internal passing line is eighty-five. If it comes in under that, the director might block the process outright.”

“Eighty-five is an ideal value, not a hard floor.” Lin Chen moved the cursor to the script’s exception-handling module. “Pull the low-confidence cases out separately. Attach screenshots of the original lab sheets and the model’s attention heat maps. Tell the director this isn’t the model being inaccurate—it’s incomplete input data. Clinical decisions are supposed to be made in context. AI is only an aid. Drawing a clear boundary is safer than propping up a pretty number.”

Su Man was silent for two seconds, then nodded. “Got it. I’ll change the output format. I’ll embed the heat maps in the PDF and mark the confidence intervals with error bars.”

The clatter of keyboards thickened again. Lin Chen lowered the tolerance threshold in the cleaning script, gave up on automatic correction of rare characters and fuzzy dates, and switched to manual verification one record at a time. Thirty. Forty. Forty-five. Each time he finished one, he drew a line on the paper checklist beside him. The paper was the back of discarded printouts, its edges already curling. Every twenty minutes or so, he had to shift his left foot—off the tower, down to the floor, then back up onto the chair bar. The numbness from poor circulation clung to his calf like a wet cloth. He couldn’t shake it off. He could only endure it.

He pulled out his phone and glanced at a bank text message. Balance: 3,421.6 yuan. The cloud server was billed by usage. If the final payment still hadn’t come through by noon tomorrow, he would have to manually scale the instance down from 4 cores and 8 gigs to 2 cores and 4 gigs. Lower specs meant slower response time, but there was no money. He would just have to bear it. He thought of the leaking roof tiles in the main room back home. Back then he had believed that if they could just fix the roof, they would finally sleep in peace. Only now did he understand: even after the roof was patched, the wind would still creep in through the cracks in the window.

At three-thirty in the morning, Su Man made two cups of instant coffee. She set the paper cups on the desk, and the steam vanished almost at once. Lin Chen took a sip. The bitterness pressed down his throat and a faint spasm rose in his stomach. He tore open a pack of soda crackers, broke one in half, and handed it to her. She took it without speaking and kept staring at the confidence curve on the screen. Between them there was no extra conversation, only the tapping of keyboards, the clicking of mice, and the occasional sound of breathing. In the third year of a startup, this kind of silence was more common than arguments. There was never enough money and always too much to do. Complaining solved nothing. All you could do was push the progress bar forward.

“Case forty-nine.” Lin Chen’s voice was flat. “The past medical history says ‘penicillin allergy,’ but amoxicillin appears in the medication record. The model treats that as a conflicting feature. Confidence drops to 0.33. Is it a doctor’s typo, or did the patient conceal something?”

“Check the original HIS log.” Su Man pulled up the interface documentation. “If the log doesn’t show any changes to the medication record, then it’s an entry error. Handle it as an error. Model output: ‘recommend review of medication history.’”

“Okay.” Lin Chen added a conditional branch to the script. if allergy_conflict: flag='manual_review', confidence=0.35

He was not chasing one hundred percent automation. Real-world data was never clean, and models were not gods. Mark the risks. Clarify the responsibility. That was the basic rule for turning technology into something usable. It was an understanding he had bought with three years of trial and error, and one line in his notebook of mistakes that he had underlined again and again.

At five in the morning, a gray-white light began to spread outside the window. Rain slanted against the glass with a soft, granular hiss. The fiftieth record was verified. Lin Chen imported the cleaned spreadsheet into the inference engine and generated the final report. The PDF was 14.2 MB. Every page carried a timestamp and a hash verification code. He checked the formatting three times to make sure there was no garbled text, no unauthorized fields, no conclusions that promised more than they should.

“The draft is done.” He dragged the file into an encrypted archive, set a password, and sent it to Xiao Li in the procurement office by email, copying Su Man.

Su Man leaned back in her chair and let out a long breath. She rubbed between her brows, her voice very light. “Can we sleep for two hours?”

“You sleep first.” Lin Chen opened his notebook of mistakes to a fresh page, lowered the tip of his pen, and wrote:

“Entry 235: Real-world data validation. Risk: uneven quality of raw data, confidence fluctuations caused by missing-value handling, strong subjectivity in clinical alignment standards. Countermeasure: abandon full automatic imputation, preserve a manual review channel, output auxiliary conclusions with confidence intervals. Execution: report delivered, awaiting feedback from the hospital affairs meeting. Do not press. Do not explain. Only prepare the original logs.”

He closed the notebook and turned off the main monitor. The instant the screen went dark, the office sank into a stillness so deep it seemed almost solid. Only the low hum of the server fan remained, and the rain outside the window. He closed his eyes and leaned back. The ache in his left foot grew sharper in stillness, like a string pulled taut. He didn’t move, only adjusted his breathing. The money was almost gone. The remaining payment still hadn’t been approved. Xiaoman’s hospital deposit would last only one more week. But those thoughts stayed in his mind for only a second before he forced them down again. Anxiety solved nothing. Only the next move did.

Nine ten in the morning. His phone vibrated.

Lin Chen opened his eyes and reached for it. It wasn’t Xiao Li from procurement. It was the number of the director of the hospital’s IT department. He sat up straight and answered.

“Mr. Lin, we received the report.” The other man’s voice was steady, carrying that particular institutional cadence. “The hospital affairs meeting has been moved up to Tuesday afternoon. The director reviewed your de-identification comparison and thinks your approach is on the right track. But a PDF alone isn’t enough. Tuesday at two in the afternoon, bring your inference engine and the original logs and run it once in our internal test environment. The IT department will provide an isolated terminal with no external network access. You deploy it yourselves. You run the data yourselves. The director wants to see with his own eyes how the model goes from input to output.”

Lin Chen tightened his grip on the phone slightly.

Internal-network deployment. Isolated terminal. No internet.

That meant they could not rely on cloud computing power. They would have to complete model loading, data parsing, and real-time inference on a local standalone machine. And their core model, to save money, had always run on usage-billed cloud servers. Local deployment would mean repackaging dependencies, adapting to the hospital’s aging GPU, and completing all verification offline. They had less than forty-eight hours.

“Understood.” Lin Chen’s voice did not change. “We’ll bring the equipment over by one-thirty Tuesday afternoon.”

“Good. Don’t bring unnecessary people. The IT department’s access control is strict. The terminal will only open one debugging port, and it locks automatically if you run over time.” The man paused. “Mr. Lin, this time it’s not about procedure. It’s about whether you actually have the goods. Don’t drop the ball.”

The call ended. Lin Chen set down the phone and the screen went dark. He turned to look at Su Man. She was already awake, staring at him. There was no drowsiness in her eyes now, only alert wariness.

“Internal-network standalone deployment,” Lin Chen said. “We have to prune the model and package the dependencies for offline use. That hospital test machine is probably running an old graphics card from three years ago. It won’t have enough VRAM for the full parameter load. We’ll have to do quantization and compression.”

Su Man sat up straight, her fingers already resting on the keyboard. “INT8 quantization will cost us some accuracy, but it’ll get it under 4GB of VRAM. I can have a test version tonight. You handle the offline loading script and the log-capture module.”

“Fine.” Lin Chen pulled open the drawer, took out the last box of pain patches, tore one open, and pressed it over his ankle. A cold sensation seeped into his skin and held the pain down for the moment. He opened a terminal and created a new project folder.

Name: hospital_demo_v1_local

The rain had stopped outside. A split opened in the clouds, and sunlight cut slantwise into the office, falling across the desktop and lighting the dust suspended in the air. Lin Chen typed the first import command. The progress bar had not yet reached the end, but the road was already laid beneath his feet. Tuesday at two in the afternoon. Isolated terminal. Offline environment. One shot to decide it all. He knew that compliance was only the entry ticket. Efficacy was the real trump card. And trump cards had to be played one by one, by his own hand.

On the terminal screen, the cursor blinked quietly, like a star gathering force somewhere in the dark. He drew a deep breath and let his fingertips fall. The next line of code was already there.