Chapter 272: Noise and Signal

The cold light of the screen fell across Lin Chen's face. The ward heating had long since shut off, and the air held the mixed smell of disinfectant and old bedding. He drew his hands back from the keyboard; his knuckles were stiff with cold. The spasm in his left ankle peaked at two in the morning, the muscle tightening again and again as if twisted by a thin wire. He did not reach for the medicine. He only tucked a hot-water bag into his pant leg, leaned back in the chair, and waited for the dull pain to ebb.

The data-cleaning pipeline was already in place. The 3.2 GB of raw files from Municipal Second Hospital had been broken into CSV chunks sorted by patient ID and timestamp. The "atypical seizure cohort" mentioned in Director Zhao's paper did not exist as an independent label in the raw data; it could only be reconstructed from vague phrases in the clinical notes, terms like "suspected focal origin" and "with autonomic symptoms." Lin Chen wrote a regular-expression matching script to align keywords in the medical records with the time windows of the EEG segments. After the first pass, it filtered out 412 suspected records.

He opened the signal-processing module. The 8-12 Hz microtremor from muscle activity was normally smoothed away as noise by the high-pass filter, but Director Zhao's paper noted that this band of low-frequency energy would show a phase shift seventy-two hours before a seizure. Lin Chen did not apply an off-the-shelf deep-learning model directly; edge devices could not bear the computation or power draw. Instead he used a discrete wavelet transform, decomposing the signal into different scales, preserving low-frequency details, and then calculating energy entropy with a sliding window. Line after line of code went in, and the log file filled with dense debugging messages.

At four in the morning, the first round of feature extraction finished. The confidence threshold was set at 0.80. Out of the 412 data segments, the algorithm flagged 87 warning signals. He compared them one by one against the clinical records: 61 matched the timing of later major seizures, with a lead time of three to six minutes; in the remaining 26 cases, the clinical notes showed that the patients had merely turned over or coughed, but the algorithm had still raised false positives.

The false-alarm rate was too high. Lin Chen stared at the screen, his fingers tapping unconsciously on the desktop. In a clinical setting, a false alarm was not a number. It was a nurse being woken in irritation at midnight, a family member's vigilance gradually going numb, the final outcome of a device being unplugged. He had to push the false alarms down, but he could not sacrifice that precious four-minute warning window.

He reopened the filter parameters. The notch frequency for power-line interference was adjusted from 50 Hz to 49.5 Hz, avoiding part of the harmonic from some patients' baseline heart rates. The sliding-window length was stretched from two seconds to three, trading a little response speed for a more stable energy baseline. The threshold was lowered from 0.80 to 0.76, but a hard rule was added: only three consecutive triggered windows would output a warning. Once the code was changed, he ran the batch again. The progress bar crept forward slowly, and the hum of the CPU fan sounded especially clear in the silent ward.

At half past seven in the morning, Su Man pushed the door open. She had not brought breakfast, only a bag of cooling patches and two cans of black coffee. She tossed the patches onto the table, pulled over a chair, and sat down, her eyes going straight to the log on the screen.

"How far has the false-alarm rate dropped?"

"0.38. But the miss rate rose by 0.05." Lin Chen handed her a coffee, his voice a little hoarse. "Director Zhao wants usability, not a perfect lab curve. Nurses can still accept 0.38 false alarms. Push it any lower, and the warning window breaks apart."

Su Man opened the coffee, took a sip, and frowned slightly. "I ran the low-temperature curve for the power module again last night. At minus five degrees, the voltage drop triggers a second frequency reduction. If the corridor temperature is low on demonstration day, the sampling rate may fall to 80 Hz. Will your wavelet transform distort at 80 Hz?"

Lin Chen froze for a second. He immediately opened the simulation script for sampling-rate reduction. He entered the 80 Hz parameter and reran feature extraction. The waveform on the screen became rough, the low-frequency details were shaved down, and the energy-entropy calculation began to jump. The confidence score fell straight from 0.76 to 0.61.

"It will," Lin Chen said. "We need to add a linear-interpolation compensation step and fill the missing sampling points before the downshift triggers. But it will add 0.1 seconds of computational delay."

"Add it." Su Man did not hesitate. "What the clinic needs is uninterrupted power. A 0.1-second delay is acceptable. Hard-code the compensation logic into the firmware. Don't leave a dynamic switch. During the demo, the more fixed the parameters are, the more controllable Director Zhao will feel it is."

Lin Chen nodded, and his fingers returned to the keyboard. The interpolation module was not complicated, but it had to connect seamlessly with the existing buffer queue. He checked the boundary conditions line by line to prevent an array overflow. After writing it, he compiled and flashed the firmware. The engineering unit's indicator light blinked twice and returned to standby.

At ten in the morning, Xiao Man woke up. His eyes were still a little unfocused, but he recognized Lin Chen. Lin Chen handed him warm water and watched him slowly finish drinking it. The monitor's beeping remained steady, with no abnormal waveform. Lin Chen connected the spare lead wires to the engineering unit and ran a real-time monitoring pass. The confidence curve on the screen fluctuated smoothly between 0.72 and 0.78, never crossing the line.

That afternoon, Lin Chen began preparing the demonstration materials. He deleted every algorithm-architecture diagram from the PowerPoint and replaced them with three clinical-path comparison tables: the response time of traditional monitoring, the battery-life shortcomings of existing devices, and V1.2's warning-trigger logic in downshift mode. Each page kept only the core conclusion, not the derivation process. He knew Director Zhao's time was expensive, and his patience even more so. The mistake startups made most easily was treating technical details as their moat. In a doctor's eyes, there were only two moats: do not make trouble, and be able to catch what falls.

At eight that night, the foot injury struck back again. Lin Chen soaked a towel in hot water and laid it over his ankle. The twitching of the muscle traveled from under the skin into the bone. He bit down on the towel and made no sound. Xiao Man slept heavily in the next bed, breathing evenly. Lin Chen closed his eyes and ran through tomorrow's route in his head: arrive at the teaching room forty minutes early, check the projector connector, test the engineering unit at full charge, prepare two spare lead wires. If Director Zhao asked about data compliance, show the ethics approval and the de-identification protocol directly. Promise no cure, only assistance. Do not compete on parameters; compete on stability.

At two in the morning, the final pressure test finished. Across the 412 atypical data segments, the algorithm output warning signals 79 times. After clinical review, there were 64 true seizures and 15 false alarms. The false-alarm rate was 0.37, the miss rate 0.04. The average lead time was 4 minutes and 11 seconds.

Lin Chen locked the parameters. He changed nothing more. He knew that if he kept tuning, he would only fall into the trap of overfitting. Engineering was not mathematics. There was no absolute optimal solution, only the compromise most suitable for the moment.

At 8:20 on Wednesday morning, Lin Chen arrived early at the neurology teaching room. The corridor was very quiet; only the rubber wheels of the cleaner's cart sounded as an auntie pushed it past. He connected the engineering unit to the projector, checked the power, and tested the lead wires. The screen lit up. Self-check passed. Voltage: 3.52 V. Sampling rate: 100 Hz. Warning module: standby.

At 8:45, the door opened. Director Zhao walked in. He carried no medical-record folder, only a kraft-paper document envelope. He was in his early fifties, with gray-white hair and very calm eyes. Two residents followed behind him.

"You're Lin Chen, right." Director Zhao set the envelope on the table without looking at the projection screen or asking about the device parameters. "No need to go through the process. I've read your company's TS terms. Capital wants returns. I want patients to get one whole night's sleep."

He opened the envelope, drew out three freshly printed EEG records, and pushed them in front of Lin Chen. The waveforms on the paper still held the warmth of the printer.

"These are real-time records from three new patients admitted last night. The labels aren't done, and the artifacts aren't cleaned." Director Zhao looked at him, his tone level. "Connect the leads now and run your algorithm. If it can warn before their seizures, the filing can be signed today. If it can't, or if it gives more than three false alarms, then your product stays in the lab."

Lin Chen lowered his head and looked at the three waveform charts. The signals on the paper were messy, the baseline drift severe, and muscle artifacts crowded the lines like burrs. His fingers tightened slightly, and a familiar stab of pain came from his left foot. He said nothing. He only pushed the engineering unit to the edge of the table and picked up the lead wires.

The progress bar on the screen began to move.